That's quite a neat way of starting. Of course the fatality rate doesn't reduce linearly with the size of the the available cohort (it selects the most vulnerable first), but as this would only apply to an immediate homogeneous geographical spread (so that it has access to the entire most vulnerable cohort at once) I think the two factors might balance one another out.. ish.

What's nice about the way you've presented it is that it shows how the short term predictions (the ones justifying the responses) aren't affected by the exhaustion of cohorts. They don't need to take it into account because ICUs are overstretched before the effect even kicks in.

I take it from this that overwhelming emotions, and at least part of classical view, remains intact. — Luke

No. The classical theory is that stimuli put the brain into a recognisable state (which can be labelled a particular emotion) which then either causes, or makes more likely, a particular set of responses. It's a one way process and the recognition of that state (by either yourself or an outside observer) is simply and act of journalism, just noting what is the case.

The trouble is there's no evidence that this is the case. None of the states we talk about (anger, fear, jealousy...) can be recognised physiologically, we do not reliably report the same physiological states in each of these categories and neither do outside observers.

Barrett's theory is that the emotional state of the brain is no less a part of the active inference model of cognitive process than, say, perception is. Stimuli put a part of our brain into a certain state. Higher order parts of the brain then try to predict the reason why these lower order parts are in the state they're in. That prediction acts as both a forward-acting imperative (creates behaviour) and a backward-acting suppressor (tunes out conflicting data). The forward-acting imperative is some behaviour solely designed to help this part of the brain confirm or deny it's prediction. This process then continues with the lower part of the brain now being put into a new state resulting from stimuli caused by the actions the higher part of the brain just initiated.

(I'm using lower and higher here as hierarchical terms, it's nothing to do with animal/rational, or basic/advanced as these terms are often used to mean)

So emotion is a model of some higher order part of the brain to explain the state of several lower order parts of the brain in response to stimuli, then to initiate some action to both filter results assuming that model is the case and to interact with the environment in such a way as to confirm that model.

This explains why there's no strictly applicable physiological signatures to our emotions, one model only need be sufficient to predict a cause of the stimuli and initiate action to better predict that cause, it need not act as a filing system sorting those causes into reliable categories.

The 'labelling' of these stimuli as being in an emotional category, is itself one of the actions the second order part of the brain is carrying out to either firm-up or cast aside its inference model. The labelling itself acts as a filter/action initiator alongside other actions.

I've oversimplified that a lot just to try and get a short overview, hopefully the main point still carries.

The ability to control our emotions in some ways is a side-effect of the fact that third order parts of our brain are also using active inference models to predict why those second order parts are in the state they're in (have the models they have). we can use the backward-acting responses from these to interfere with or constrain the models they choose.

When we can't control our emotions is then a matter of there being no access (no circuitry) between the second order part of the brain and the actions it initiates. It doesn't have any impact on Barrett's theory which is about the indeterminacy of emotional states (and the reasons why they are indeterminate).

For involuntary emotional responses the two theories would look like this;

Classic - Stimuli (I stub my toe) > emotion (anger) > unavoidable response (I yell obscenities)

Active Inference - Stimuli (I stub my toes) > several lower order neural circuits are put into various states (pain, adrenal response, muscle contractions...) > a second order circuit uses the model 'anger' to predict the cause of the states of all these lower order circuits (some external threat is hurting me) and sends out action initiators to yell obscenities(frighten off the external state) > no third order circuit interferes with these action initiators (I yell obscenities) > the first order circuits report the response of the environment to my actions (the pain stopped) > the second order circuit either updates its model according the amount of errors deviating from its expectations, or suppresses deviant information (all good - I yelled at the cause of pain and the pain went away - well done me).

You might have already read this one already given your interest, but reading you two talk about language (I'm afraid much of which is going over my head) reminded me that Barrett did write a paper on language and emotion which might touch on some of what you're thinking about (or possibly be completely unrelated, but you might find it interesting nonetheless).

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2225544/#!po=37.6923

I was thinking Roman army divisions - not exhausted even by decimation. — unenlightened

Yeah, this seems a common misconception. That because the fatality rate is 1% it kills 1% of any population it's exposed to, as if the virus itself has a quota to fill. The 1% is a feature of the population, not the virus. It's saying we, as a population, are in a state of risk distribution such that 1% of us will be killed if exposed. Once that 1% has been killed we are no longer the same population, we no longer have the same risk distribution.

As a (socio-political) aside. It's interesting how little focus there is on the fact that the fatality rate is a measure of the health of our population (and by association, the quality of our healthcare). The same people who are decrying the lamentable state of our healthcare system in its (in)ability to respond to this crisis seem (to me) to be the same people wanting desperately to downplay the relationship between poor health and increased Covid-19 mortality which that same lamentable state is directly responsible for. But maybe I'm reading the wrong people.

No luck finding the analysis I'm afraid. Shame because I'd be quite interested (albeit only academically), but to do it from multiple RRs even if I had the prevalence data for those groups would require both the computer and the statistician (I'm not good enough to do my own stats) from work. As I'm now (re)retired, that would be quite a stretch for an idle speculation.

One way to look at the final figure though is the method I mentioned to boethius using a cohort of {those with underlying conditions serious enough to be listed as a cause of death}. We know Covid-19 mortality comes almost entirely from this group (91-98% in the reports I've read), and those with two or more have at least triple the RR, so they contribute more than their numerical share to the mortality.

So, once this group is exhausted, CFR will drop to at or below the CFR for the less affected groups (less than 0.1).

The US death rate is about 2.9 million per year (and almost half of those are accidents or intentional self-harm) , so even if Covid-19 attacked every single one it would be difficult to reach your target without exhausting the group from which almost all fatalities are drawn.

I don't think we'd have the data. What I'd need is the CFR (or better IFR) for a stratified set of cohorts. What we have from the Lancet study is the answer to the question "of everyone who caught this disease and died, how many were 70-80?" (or whatever cohort size). What I'd need is the answer to the question "of all the people who were 70-80 who caught the disease, how many died of it?". I don't think anyone has done that yet.

Another possibility is to use the RR values for the prognostic factors, if I could find data on the prevelence of those factors (they're so important for loads of conditions, I expect that data is out there). I'll have a look and see what I can find.

To clarify though, with every country in some form of lockdown these numbers are purely speculative. The r values are going to drop to too low a rate before any major age cohort is exhausted.

It looks to me as though that timescale is about up to the point where herd immunity might become a factor. Cohorts will not be exhausted as long as the virus is spreading geographically to new populations. Is that right? — unenlightened

Cohorts are technically exhausted the moment one person dies, the cohort {most likely to die from condition x} is fully exhausted as soon as someone dies from condition x. Depends on the specificity of the cohort. The effect is that the make-up of (and therefore the risk distribution within) a cohort will change depending on the variables it is exposed to.

Geographic spread could affect the rate at which cohorts are exhausted (one localised sub-section of a cohort might become fully exhausted before the disease has spread to the next), it would also affect the rate of increase if the cohorts are not geographically homogeneous, but I don't think either of those factors will affect things on a national scale - maybe though. I'm sure some states/countries have a different age distribution and so fatality rates would rise/fall as the disease reaches those areas.

I got it from here : https://www.thelancet.com/journals/laninf/article/PIIS1473-3099(20)30243-7/fulltext

I'd say it's the best estimate we have so far. — Benkei

It's not the quality of the estimate that's the problem, it's extrapolating it to changing cohorts. Image a 100% infectious disease only killed men (but did so every time). At first the CFR for society would be 50% (for every 100 infected 50 died) , but it could not sustain that ratio as within time (depending on R values) there would be twice as many women as men. Killing all infected men would still only yield a societal CFR of 25%.

Thankfully (because of the lockdowns) we're unlikely to be dealing with total deaths anywhere near big enough for the effect of changing the risk profile within cohorts to be so large, but it is incorrect to use these unadjusted figures to imply such massive numbers as your model does in the case of an uncontrolled spread.

In expectation terms the uncontrolled spread would kill 90% of the cohort who are at 90% risk, 80% of the cohort at 80% risk... And so on, with those risks being calculated independently (ie from within their cohorts).

Once it's killed 90% of the 90% risk cohort, it's not going to start killing more in the less at risk group to 'make up the numbers'.

Doing absolutely nothing will mean your ICUs are overloaded in week 18 assuming they all have ventilators. The next week you run out of enough beds to take care of hospitalised infected. Somewhere in week 22 you will have over 40% infected and herd immunity will slow the spread. I don't know how much, so I haven't taken it into account for the two weeks thereafter (so you should ignore those). By week 22 almost 2,9 million US citizens will have died (actually, that number is probably delayed by a couple of weeks). — Benkei

Your estimate is based on a misunderstanding of the statistics. 0.6% is not a target. It doesn't act as some kind of quota the virus is trying to fill. It's a summary of the frequencies which have been observed so far, all of which reflect the combined action of underlying (hidden) variables. So, to take one such variable - d-dimer greater than 1 μg/ml. It increases the risk of mortality 18 fold. Once the cohort of people with comorbidities likely to lead to such a score has been exhausted, remaining cohorts then have 1/18th of the relative chance of dying in that variable alone. Age, organ condition and hypertension are all documented factors raising relative risk way above statistical significance. As these cohorts become exhausted the fatality rate will drop dramatically (as you can see with the example I gave, the effect on risk is not small).

You cannot accurately predict the death rate using a snapshot of the fatality rate at a given moment in time and simply extrapolate unless you use a very short timescale. You have to estimate the variables leading to death (as the experts are now doing) and produce a multi-variate model based on a declining cohort.

can't accept an epistemology of probability that won't let me say that there are probabilities I don't know. — Pneumenon

Why not? Put it like this. If you want to insert a frequentist theory of probability into your epistemology - how many trials does it take until you do know the probability? If after a thousand trials the proportion of heads to tails is 501:499 do you now know the probability is 50.1/49.9? Because if you do, I'd say you were wrong.

the more conceptual paper is her "Solving the Emotion Paradox" paper, which I think can be found with a Google search. — StreetlightX

Here, save you the trouble of typing it.

Our behaviour itself, our emotions themselves, are a sample from the model, and "exactly what we do" is a collapsed down form, a representative summary, of the model's state given (its own representation of) a current goal (and our expectations of environmental/bodily behaviours). — fdrake

Yes, that's right. The point I was trying to make above is that making any conceptual/verbal representation of this snapshot summary that we label with an emotional term is also a behaviour. So we're not just journal writing, we're not keeping a log of how we're feeling just for posterity, our drawing together a snapshot valuation of all our various emotion-related stimuli is itself an act which is part of the perceive>infer>respond>perceive(more closely) system. We're deliberately paying more attention to the contributory stimuli and deliberately trying to form a conceptual valuation of them in order to achieve some situational goal.

To put it more colloquially, we're not in some state we would term an emotion all the time (and just unaware of it). The act of terming a state and emotion is something we do in relation to some goal. Up to that point we simply have affect, no emotion at all. It's not something we discover about our state that was there all along, it's something we construct from the components of our state for some other purpose. That where the 'constructed' bit of constructed emotions comes from.

There are probabilities I don't know. The probability of a coin landing on heads is not exactly 50/50, but I don't know what the exact ratio is. — Pneumenon

You may not know what the exact ratio is, but that would only matter in frequentist probability. Your mental models appear to be better explained by Bayesian probability. So the ratio doesn't matter. If you'd bet no more on heads than you would on tails then the probability is 50/50. Your belief is that it is no more likely to land on heads than it is on tails.

It might help to think of it as two questions about your knowledge (priors).

Do you know something about heads which biases the coin in its favour? Do you know something about tails which biases the coin in its favour?

An answer of "I don't know" doesn't make any sense here because it's not a question about the world, it's a question about your knowledge of it.

Since you have no more reason to think it will land heads than you do to think it will land tails, the probability is 50/50.

The presence of some piece of data in the brain is not implied by asking you either of those things, as far as I can see. — Pneumenon

No, you're right, I was reading too much in to your comment. Most people, when they talk about belief, refer to it that way - "Jim believes in God", "Bob believes in UFOs"... As if there were some part of their brain in either an on or an off state for each posit.

Our brains just don't seem to work like that. Our beliefs are probabilities, not switches and they're tested and updated all the time depending on the stimuli we're trying to model. Your disposition toward God or UFOs or even where your front door is will not be the same tomorrow as it is today.

2.

I think this...

I imagine there are other nuances of the errors we can make in feeling; — fdrake

... is important to understand the implications of this model. One experiment done on generating responses in mice introduced an element of randomness to see if the inference of valence to a Pavlovian response would still confirm to a purely Bayesian model (bit of background, it does without valence, classic experiment on correcting errors in sensory conflict showed the predictions were almost perfectly Bayesian). On this occasion, they didn't. As soon as the expectation had valence the predictive model erred from purely probabilistic. Basically, the mice were reluctant to update their priors to reflect the probabilities they were experiencing when the expectation had valence. The model actually included the valuation of the result.

So with modelling emotion, we're not necessarily just modelling the most likely cause of the stimuli (and appropriate response), we're biasing those models in favour of certain predictions depending on the value we previously gave them. Technically an error. We act as if a particular model is a better explanation than it actually is.

... Barrett's point about cognitive systems maintaining allostasis is, I think, more than just a neuroscientific one. It emphasises the role of interoception as forward-acting (it generates some following set of reactions, rather than just backward-acting, suppressing existing responses) something else afterwards . So a model, is a kind "this is like/is due to/arises from that" (to borrow @fdrake's expression), but it cannot really be separated from "...and this is what I do about it" because the doing is part of the loop updating the priors.

A couple of things that might be useful to bear in mind with this stuff.

1.

Our valuation of the interoception/perception of stimuli is not idle journal-writing. The prediction the we vocalise as an emotion category is the attempt to render into language a model of our state which actually has a purpose beyond that report. In the classical model this is already taken care of (the 'emotion' puts the brain in a state better able to carry out the task at hand). With an active inference model though, we have a much more interesting intersection. So our model predicts the cause of our state, but, as with perception, it's a proactive model, it tests the theory by taking action 'as if' it were the case and responding to errors.

In perception, this might take the form of looking for edges or forms we expect to be there (once we've predicted it's a rabbit, we look for the ears).

In affect modelling, we'd be doing something like focusing on our skin response once we've formed a predictive model of fear based on, say, our heart rate. This can extend to external responses too, so that aspects of our environment become brought into focus depending on their role in the whole 'anger' story-line. The emotion is not just felt within our own bodies but is an interactive experience with our environment. Others take part in it.

Evidence for this comes from differential emotional reports in different environments in response to the same stimuli.

(Took me longer to write that than I thought it would... 2. later...)

Now we are seeing the brainwashed people coming out to protest in the US. — Punshhh

I'm always wary of assigning positions to 'brainwashing'. Not because it's not appropriate, but because I don't think it's helpful. The mechanisms behind brainwashing are present in literally every thought you, I, or anyone else has. It's not a binomial state, it's about degree. The "we must all stay indoors to help fight this global crisis" is no less a narrative than the "it's all a global conspiracy" (for what, we don't yet know!). Just because the former is more true, doesn't make it less of a narrative. That's important in a situation where the state of scientific knowledge is changing rapidly. People will update their narratives much more slowly, and no less so if they were right in the first instance.

You're right about the signal that's driving this, but with 7 million premature deaths linked to air pollution, the same could be said of anyone driving their car into the town centre. With 1.9 million deaths from diarrhoeal diseases directly related to poverty, the same could be said of anyone not paying a fair price for agricultural products from developing countries. It comes down to beliefs about the weight of responsibility vs autonomy. We're more forgiving of slight variations in that balance when we have more data (it's easier for us to see complexity in larger datasets). So here, battle lines are more stark because the dataset is small.

Trump is trying to present himself as a stable genius who has got a handle on this virus and is taking all the right actions and responses. — Punshhh

Yes, it's laughable isn't it. Of all the characters he could have potentially got away with presenting, stable genius was not a good choice. Mad-max-like anti-hero might have worked, stable genius is a reach even for such a consummate liar as he is.

Here in the UK we have a curious juxtaposition between the populism and a sense of civil obedience and cooperation. — Punshhh

Is it such a juxtaposition though? I see what you mean, but the responsible media (and even scientists) are not made up of people magically immune from influence by their social groups. We shouldn't mistake the clear boundaries to reasonable belief created by science for a guide to 'right' belief. It's not the same thing at all.

One thing that's interesting for me with this crisis (this thread being a good example) is the narrowness of ideological branding. I'm not getting into the conspiracy bullshit, I'm meaning within the parameters of what is scientifically valid opinion, certain positions are being allocated to political ideologies to which I don't think they belong. I don't believe there's such a thing as a non-political view. All views come from underlying ideologies which have political ramifications. With Coronavirus there's variables - the extent to which it's a crisis, the proportion who will be affected, the effectiveness of certain strategies, the cost/benefit of certain strategies. In less critical times, there might be a range of each of these variables associated with the range of political ideologies (whatever your favourite two-axis compass). Here I feel there's a basic association of all valences with either right or left. Back to the impoverished understanding we had of political spectra before Eysenck even. Can you distinguish a left-libertarian version of this from a left-authoritarian version? Or the libertarian capitalist from the state capitalist response? It seems much more right-wing/left-wing and no second (or third) axis.

as their populist message became superseded by a global pandemic the populism has become curiously silent and the population has fallen into line behind the instructions of the medical experts. — Punshhh

Yes, this is an interesting phenomena. I had a colleague at work who would divide the religious into those who believed in God and those who BELIVED in God. The latter group, he said, were identifiable becasue they acted as if the Devil were literally behind them with the red hot poker ready to insert. The former group would change Gods if it offered them a better deal at the supermarket. The point is that I think feeling one's life (or those of ones close social group) is at risk really undercuts beliefs which were held only for convenience, but it does not dent those which were held fundamentally. I guess America has more fundamentalists.

I might prefer to say something like: there are many states called anger, each of them variantly evoked and produced across a range of different situations. — StreetlightX

Yes, that's a better way of putting it from our (phenomenological?, never sure how to use that word properly) perspective.

From a neuroscience perspective, I think the current thinking (Sapolsky, Seth, LeDoux...) is that the collection of these states has no (or little) neurological significance, as a group. By which I mean constituent states (affects, perceptions... ) which form part of one of the experienced states called 'anger' have no more connections with each other than they do with constituent states typically associated with other emotional classes. Does that make sense?

I realise that position doesn’t really have much use here where we're talking about the consequences for us as we experience these things though. Just thought it might help underline the force behind Barrett's position.



It's not really philosophy. In order to be useful a class has to be such that it's members have some usefully distinguishing characteristics that are not shared in equal significance with similar members of other classes.

All that's happened here is we started investigating models of brain function using the classes given to us by our experience and we found them to be not so useful because we couldn't isolate any characteristics which uniquely identify members of that class. I'm not sure how they might better have proceeded, given the evidence they had.

Barrett's theory has absolutely nothing to do with the mere act of labelling sensations. It's not about saying "this is not 'anger', this is 'anger'" - She's arguing, in her 2006 paper, (very convincingly from an evidential point of view) that there is no such state as 'anger'. If you looked for evidence of such a state in the brain you will fail to find any such thing. Same goes for all the other emotions. Likewise if you look for such a state psychologically, all you will find is a collection of behaviours no core of which can be demonstrated to be necessary, no single one of which sufficient.

In her 2017 paper she uses modern theories of computational neuroscience (active inference) to posit a possible means by which we come to categorise certain disparate and dissimilar collections of perception and interoception in the same class.

As @StreetlightX says in the OP, this has many interesting implications - it does for psychology no less, but the matter of whether she's right in the first instance shouldn't really be up for debate here. The question she's answering is the question of how we come to categorise disparate sensations in the same class. The fact the we do this from a neuroscientific point of view is relatively indisputable (at least not disputable without reference to neurological evidence of emotional states having some unified and identifiable correlates in the brain).

Her model might be wrong, but the need for an explanatory model is a live issue in both neuroscience and psychology, it's not a superficial re-branding.

We don't have to assume 50/50, do we? Couldn't one start at assuming a 100% probability of heads, leaving it to the Bayesian process to level out at 50/50? I'm just wondering what the correct process is. — Banno

Yeah. I'm no expert in Bayesian mathematics, so I couldn't say if the actual function is expected to carried out that way, but insofar as it applies to cognitive processes, the brain is remarkably confident about supplying us with priors and it will do so with only the slightest provocation (see sensory isolation experiments), so you're, right, it probably is more accurate to say we start off with something other than 50/50 - some past experience getting lucky with heads, some half-considered notion that heads are heavier, maybe even (as very young children) some simple misconception about how physics works. Something probably will bias our initial belief and we revise it down as we get an increasing number of errors using our 70/30 or 100/0 models.

You can see how supernatural beliefs get maintained this way though. By the time you've revised your model down to say 60/40, you're getting increasingly less likely to actually notice the errors in your model unless you really experiment in the abstract. Someone who had a prior belief that a coin was 60% likely to land heads and 40% likely to land tails (because 'God prefers heads' or whatever), will have little reason to update that prior unless they actually toss a thousand coins and plot the results (and believe in maths, of course)

Perhaps you can answer: I don't know the probability of the dolphin. This is admissible — Pneumenon

Not with Bayesian probability it isn't. You don't know if the coin will land on heads or tails. That makes the probability 50/50.

The error here is assuming that belief is some identifiable bit of data lodged in the brain somewhere, that of any given object (once described to me) I will create a byte of data corresponding binomially to whether I believe it or not (see ). There is very little evidence to support this theory of belief. Rather beliefs are dispositional states. My belief that there is an invisible miniature dolphin swimming in continuous circles around my head is a measure of the extent to which I'm going to act as if that were the case. This allows for (in fact necessitates) grades of belief - I can easily act as if it might rain today without having to form a belief that it will or won't.

Not only is belief graded this way, but it is constantly updated. The evidence from our perception becomes the prior for our next inference, so we rarely have a consistent state of belief about any given thing unless we're interacting with it, and even then our belief is changing all the while we're manipulating the object (works for concepts too).

Furthermore, our beliefs then affect our perception in turn. So a disposition to believe what I see is a tree will lead me to actively look for leaves, a trunk, branches... the things I'm expecting to see there because it's a tree. I will disregard parts that don't conform to 'tree'. Again, the same works for concepts.

So active disbelief is only really a verbal state with 99% of examples. to actively disbelieve the teapot is only to say that if someone asked me if there was a teapot orbiting mars I should be inclined to say no. If no-one asks I have no belief state about it at all. Not zero, just no state.

Likewise a belief in some object (or entity) is not the same for each person. To say "I believe in God" does not mean the same thing from the devout fanatic as it does from the social church-goer because the resulting behaviours are different.

@Punshhh - should have tagged you in to the post above

such a contrary contention would require a substantial amount medical evidence demonstrating its veracity and the mechanism by which it acts. — Isaac

I should also add to this that I'm speaking hypothetically. This work has already been done and disputes the claim.

Independent predictors of mortality were diabetes mellitus, a history of renal dysfunction (or higher creatinine), New York Heart Association (NYHA) functional class III or IV, lower weight or body mass index, lower blood pressure, ankle oedema, and higher scores on a disease specific quality of life questionnaire...A prognostic model produced on the basis of easily obtainable information from medical history and physical examination can adequately stratify heart failure patients according to their short term risk of death. (my bolding) — From the BMJ

It is not a mystery what factors are linked to mortality in people with serious conditions, there are entire libraries filled with papers about prognosis of mortality from various conditions.

Thanks, that ties in with what I thought you were doing, but I wasn't sure.

So it might help to put some numbers in?

Group 1 - those who are going to die this year is about 500,000, but when we're talking about overlap of comorbidities, we're only really interested in a sub-group {those who are going to die this year from underlying health issues}. That's about 300,000 - taking away accident and intentional self-harm.

Group 2 - comorbidities which will not lead to death this year. About 2.5 million for cancer, 7.4 million for heart disease. Other risk groups are much smaller, so we could say about 11 million.

Group 3 - some proportion of the remainder (about 59 million) who will, despite a lack of comorbidity die from coronavirus. We know from the studies that this group is somewhere between 0 and 9 % of all coronavirus deaths, so taking Prof Ferguson's latest estimate of 20,000, and a mid-range estimate, this group would be about 1,000 people.

So the question is how the remaining 19,000 estimated deaths will be distributed between groups 1 and 2.

We know that this group (the 19,000) will have comorbidities serious enough to be listed as a cause of death. So we can re-label this group, group A {those with comorbidities serious enough to appear as a cause of death}

As you can see, the size of groups 1 and 2 is irrelevant right now. The question is solely about the nature of group A. Is group A drawn mostly from group 1 or mostly from group 2? Group 2 being bigger only makes a difference if group A is being drawn from the pooled group 1+2 at a bias that is significantly less than the 3 in 100 ratio between the group sizes (boethius's contention).

Disputing even the lower of the estimates for overlap (50%), we'd need to argue that fewer than 49% of those in group A are drawn from group 1. Ie we'd have to say that the group of people so ill with a condition that it is listed as a cause of death is not even majoratatively drawn from a group of people so ill with that same condition that they are going to die of it later this year.

To me, such a contrary contention would require a substantial amount medical evidence demonstrating its veracity and the mechanism by which it acts. (not to mention the reason why the country's leading expert on pandemics has somehow missed this fact in his training thus far)

I'm trying to follow your line of argument here (or rather your request for clarification), but your terminology is a little confusing in places. It may just be that you're attempting to reflect boethius's terminology, but I may also have just misunderstood what you're saying, so...

If you're accepting that there could be a "large number of deaths in group 2", then group 1 cannot possibly be "small". Experts predict about 20,000 deaths total from Covid-19. Group 1 has at least 300,000 in it. Or is that the point you were trying to make and I missed it?

"End of their lives" as in over 60? — boethius

No! Who the hell thinks people over 60 are at the end of their lives. I bloody hope not.

Or, "end of their lives" as in will die within 1 year? — boethius

Yes. In the context (and supported by David Spiegelhalter, who specifically referred to 2020). I'm quite confident "end of their lives" meant they they were close enough to death to fit mostly in the year's mortality. Coupled with the severity of a comorbidity appearing as a cause of death. If someone had lung cancer recorded as the cause of death, but then (imaginary doctor incompetence) it turned out they weren't dead after all, just unconscious, do you really think their not very much less likely than other lung cancer patients to make it through the year? "It was nothing, just a little lung failure severe enough to be listed as a cause of death... I got better"

show where this expert clarified their meaning of "end of their lives" as to mean "would have died within 1 year". Otherwise, again, you are citing evidence that supports my position, not yours. — boethius

What? If I can't cite evidence he meant within exactly one year then that somehow counts as evidence supporting your position? How on earth does that work? If I can't cite such evidence (notwithstanding my other supporting evidence) then at best that means we don't know. Under no circumstances does it mean that this cohort are definitely not expected to die anyway within the year. How does it support your position?

Second thoughts just don't bother answering. I've had enough of this.

Maybe you could flesh out how you're using "random" and "caused." Random stuff is usually understood to be caused. — frank

Yeah, fair enough. I'm using random in the sense of not possible to control for. As in, some as yet hidden factor, some non-measurable element of chance (such as replication error in cellular growth), or some ubiquitous factor.

Things would have been worse if this happened 100 years ago. — frank

Interesting thought. Severely limited travel might have kept it in one place, lower population of elderly with comorbidities too. But lack of medical care on the other hand. Thing is, medical intervention is only saving a proportion of sufferers. Using the JAMA figures (which I know are preliminary) 14% went into care and 2.5% died. So presuming those that died went into care first, that care saved at most 85% (some survivors would have survived anyway). The first two factors only need to lower the total infected by, say, 80% or so and total number of deaths would have been lower even without modern medical care. Since the over 70 population has doubled in some countries in the last 100 years, plus most people lived and worked in one town/village...

Things would have been worse without the lockdowns. In some places it was overkill, but that's no one's fault. — frank

Yes, I think that's unarguable. They should have been sooner and accompanied by testing and tracing. We've known about the possibility of something like this for decades. It's shameful we weren't better prepared.

Those factors are also tied to a certain social setting. People will die in Honduras who wouldn't have died in the US. They'll die from dehydration, hypoxia, and septic shock. They could be in their 30s with no underlying health problems. — frank

Yeah, absolutely. I think I did mention it somewhere, but it should be made even more clear. All this only applies to the developed world. The overlapping comorbidities have a completely different cohort size in developing countries (and presumably within small, very poor groups in developed countries, I don't know). I'm still not sure about "no underlying health conditions". I'd need to see the data on that. Some people work from a default position that disease is random until some factor is proven. I tend to work from the position that it is caused until the random factor is demonstrated. It's just a different axiom, I suppose.

We reduced the mortality rate of a pandemic by collective action. — frank

Well, that's a very positive way to look at it. Not saying that's a bad thing. Personally, I'm more of a governments-too-concerned-about-public-image-to-act-in-a-calm-reassuring-and-timely-manner-could-well-have-killed-thousands kind of guy, but each to their own.

We're talking about deaths within 1 year, so talking about overlap with comorbidity in larger groups than "likely to die within 1 year" supports my position. — boethius

They are not comorbidity groups larger than "likely to die within a year". They are exactly comorbidity groups that are likely to die within a year. That's why the experts responsible are talking about overlap within that time scale.

Listing a comorbidity on a death certificate is not the equivalent of assigning a broad risk category. It's saying that the person was likely to have died from that condition had they not had Covid-19. That is literally the wording the ONS use.

Just to be abundantly clear about this the MCCD guidance states that a main listed cause of death must go "back through the sequence of events or conditions that led to death on subsequent lines, until you reach the one that started the fatal sequence. If the certificate has been completed properly, the condition on the lowest completed line of part I will have caused all of the conditions on the lines above it. This initiating condition, on the lowest line of part I will usually be selected as the underlying cause of death, following the ICD coding rules. WHO defines the underlying cause of death as “a) the disease or injury which initiated the train of morbid events leading directly to death"

And clearer still...

"The conditions mentioned in part two [not even the part we're talking about, a lesser subsidiary of it] must be known or suspected to have contributed to the death, not merely be other conditions which were present at the time."

Comorbidity on a death certificate is not the assignment into a broad risk category. It is the declaration of a very serious condition directly responsible (albeit sometimes in part) for the chain of events leading to death.

Most deaths within 1 year do not come from groups with 90% chance of death this year. — boethius

What? How does that even happen mathematically?

If statisticians put someone in a group of 1% risk of death due to heart disease this year, they are not saying that they were just too lazy to analyse further and see which of these people with heart disease have actually quite strong hearts (and so many 0.1% of dying) and which have "the weakest heart" (and so 90% of dying); they are saying "of 100 people in this group we expect 1 to be dead by the end of the year, but we don't know which one" — boethius

Evidence.

most deaths are from groups with small chance of death within the year, but they are large groups and so result in lot's of deaths. — boethius

Evidence, again.

things are no where close to predicting "who's going to die within 1 year". — boethius

Nor do they need to be. It is sufficient to see overlapping cohorts.

assuming the people who would die from heart disease this year have "the weakest heart" and the people with heart disease who die from Covid too have "the weakest heart". This is not what statisticians believe. — boethius

I've literally posted studies showing exactly that. Did you read any of them. They provide prognostic factors for deaths within broad groups (such as hypertension within the heart disease group) which accurately predict likelihood of death within that group. The same factors (in this case hypertension) are associated with a higher chance of death within the Covid-19 group. D-dimer count (18 fold increase) and SOFA scores (5 fold increase) are two more such factors.

If everyone, or most people, gets Covid, and most deaths arise within large risk-groups, then a very slight increase in chances of death due to surviving Covid can easily replenish all the risk groups to result in the same amount of deaths in absolute terms within the year. — boethius

Yes, but your premise is not true. Having a comorbidity of sufficient severity to class as a cause of death is not a "large risk-group" it is, as the country's leading expert in the field has said "people at the end of their lives".

The reason you aren't seeing higher mortality among younger people with no underlying health issues is the availability of oxygen, the ability to resuscitate with saline, antibiotics, pressor drugs, and so on. — frank

I'm not sure, though. I get how that would not be reflected in the comorbidities from the death certificates, but I don't see how that gets around the overlap in prognostic factors. Those, presumably, cover all age groups, and those affect severity as well as death (it's not like death is predicted by a different range of factors to severity). So the number of people getting to a point where they need critical medical care will still be influenced most strongly by the same factors influencing mortality.

If this is the case, then the numbers in critical care will still be heavily drawn from the numbers who would have ended up in critical care any that year due to the overlapping factors. Obviously much less so than with fatality. The critical care group will have a much greater flux than the "end of life" group. Plus complicating factors will have a greater impact because of that. I'd be interested if anyone has heard any modelling of the critical care group.

I'm not sure how it makes the cause of death not a big deal though (is that what you meant?). The fact that there's a 90% overlap with comorbidities serious enough to be listed as a cause of death is hugely significant for risk assessment.

That 90 percent of Covid deaths had comobidities, but that does not mean 90 percent will die within 1 year. Then you agree with my original position! — boethius

No. A comorbidity sufficient to be be mentioned on a death certificate is extremely likely to to cause death within the year. Doctors do not fill in death certificates with a list of "other stuff they also had", these are very serious conditions which are "mostly likely to be the underlying cause of death for a person of that age and sex had they not died from Covid-19". That is why Professor Ferguson described victims as being mostly "at the end of their lives"

So I am not saying that it doesn't mean these people will most likely die within a year. It absolutely does mean that. Having a comorbidity listed on the death certificate as a cause of death is very serious and anyone in that condition is very likely to die within the year. That is why, again, both the experts who have spoken on this matter have reached the same conclusion, and why you've not managed to produce a single expert saying anything to the contrary.

300,000 people die each year (from disease). These deaths are drawn, in the overwhelming majority, from the exact group of people who would have the comorbidities listed in the ONS figures as having a 90% overlap with Covid-19 fatality. I've supported that assertion for heart disease and cancer by providing studies of risk factors and prognosis.

For your claim to be true, there would have to be little overlap with this group.

We already know there is a massive overlap with Covid-19 fatality and these comorbidities (over 90%). We already know that there is a massive overlap in prognostic factors (I've cited the studies for you). So you'd have to present an argument which shows how, despite an overlap in prognostic factors, the 300,000 deaths this year are not largely drawn from the group of people ill enough with these comorbidities to have them recorded as a potential cause of death. This is, on the face of it, a ridiculous assertion for which you've yet to provide a shred of evidence.

You mean that regardless of overlap, long term lung damage may simply replenished those "at risk of dying within 1 year" then you agree with my original position! — boethius

No. "Complicate the figures" is not anywhere near "replenish the entire cohort". Again, there is no evidence that lung damage will cause future deaths in these numbers. This is just your speculation and needs evidence to support it.

the effect is not so large as to essentially balance out deaths over the year, or come anywhere close to that. — boethius

As I've said, take it up with the professionals who disagree with you, or present some counter-citations. Your personal 'rekon' that it won't be large doesn't amount to much on its own when contrasted with two experts who both think it will be between 50 and 100%.

No where have you presented any evidence that most people dying of Covid would die within 1 year. — boethius

Right. And nowhere have you presented any evidence that they wouldn't. Hence my point to @Benkei that a year was a bit arbitrary. Professor Ferguson talks about people at "the end of their lives" and Professor Spiegelhalter talks about a "very short time". If you want to interpret those expression as meaning much more than a year, you can, but I'll not join you. Someone with 5 or more years left being described as at "the end of their lives" is ridiculous.

Substantial for a statistician can easily mean "a small but statistically significant effect". — boethius

What about "many", does that now mean 'few'? Plus his recent comment is much clearer that he expects "there may end up being a minimal impact on overall mortality for 2020". Or does "minimal now mean 'massive'?

What evidence is there that the effect of overlap with people "who would otherwise die this year" is a big effect as opposed to a small effect? — boethius

Over 90% of people who have died of Covid-19 had comorbid conditions that were "mostly likely to be the underlying cause of death for a person of that age and sex had they not died from Covid-19". These were not the assignation of broad risk groups. These are the additional conditions the doctors considered life-threatening enough to be listed as a cause of death.

Prognosis for Covid-19 fatality is significantly worse for key factors which are identical to factors which also affect prognosis for the comorbid conditions listed. There are no non- overlapping factors listed in any of the studies.

We've been discussing the time frame of a year. — boethius

I'm aware of that. I was simply making the point that what might be a 60% overlap in a year could be a 90% overlap in two years. Picking one year is quite arbitrary (although it does cover seasonal variations, so it's pretty much the minimum time scale it makes any sense to compare over). Professor Ferguson and Professor Spiegelhalter are referring to the yearly mortality in their comments, as have I been.

since this apparently includes obesity and diabetis — Benkei

Just noticed this. The risk group (those who are significantly more at risk than average) include the overweight and those with diabetes. The comorbidities registered on death certificates (where the overlap comes from) do not include any such vague categories. They are actual causes of death. They're far less vague and use either ICD-10 or WHO cause of death categories.

What is practically significant isn't precise and is a matter of opinion. It appears to me you and boethius might be discussing opinions at this point which is why you aren't reaching agreement. — Benkei

You're right, and of course, the timescale matters. Thinking about overlap with deaths this year is a fairly arbitrary cut off point (why not the next two years or five). This is a problem with risk analysis in general and why people like Prof. Spiegelhalter tend to talk about Days of Life Lost rather than raw deaths, it's not because he doesn't care about the elderly and ill, it's just that there's no other way to account for effect of interventions statistically without skewing the results.

The overlap in factors affecting prognosis, however, is not just opinion (or rather it's the opinion of virtually every expert who's written on the subject). This overlap does affect the predictions in ways which are then beyond mere opinion. In order for the overlap to be statistically small, for example, we would have to have a lack of overlap in factors affecting prognosis to a greater extent than there is overlap. In order to sustain such a position one would have to assume that factors as yet undiscovered turn out to be so significant that they outweigh the overlapping factors already discovered. That seems quite a stretch.

What we know is that the vast majority of fatalities (over 90%) had other comorbidities which were "mostly likely to be the underlying cause of death for a person of that age and sex had they not died from COVID-19". so this is referring to cause of death at the time of death. Not cause of death eventually, or some time in the distant future if they're unlucky enough. It is the other factor which the doctor or coroner thought serious enough to contribute to the actual death at the time ie without Covid-19 they would quite likely have died from that condition alone.

I agree that the complicating factors of system overloading and long-term lung damage make the figures difficult to say with certainty, but there is not any evidential support for the position that the overlap with those who would have died anyway will be statistically very small. As Professor Ferguson says, this is primarily a condition which causes death in those who are already very ill.

You said...

This will lead some people to imagine there was never a problem in the first place. — I like sushi

You can't say that they'd be imagining it without having done the analysis. Presuming here we're talking somewhat rhetorically. If you literally mean people imagining there was no problem - zero problem - then of course they're wrong already, but if that's what you mean, then you're straw-manning. No one is claiming there's no problem, even the worst right-wing rats are admitting that a problem exists.

Please do stick around to discuss this, your contributions are valued. — Punshhh

Thank you. That's kind of you to say so.

I am not sure what you position is? — Punshhh

Not too far from you it seems. My main concern is the psychological impact in two major ways.

1. We needed to have responded to this crisis much quicker and with more decisiveness - let's be absolutely clear, despite my efforts to explain the overlap in deaths from other conditions, even if the overlap was 100%, having a year's worth of deaths in the space of a few weeks is an absolute disaster and would undoubtedly have caused thousands (if not tens of thousands) of unnecessary deaths due to the overloading of the healthcare systems. We needed to have instigated social distancing, testing and tracing straight away and the fact that we didn't is bordering on criminal. The problem, psychologically, is that the more the threat is hyped up, the more people panic about it, the less rationally they respond and that is the opposite of what we need. It may be tempting to think that presenting the worst case scenario fires people into action, but the literature just does not support that position. People become either hyper or hypo aroused to the threat meaning that they will either see it everywhere (and so not focus on where it really is) or they will just 'block it out' because it's too big to handle. Both of these effects are well-documented (it's not just guesswork) and both of them could be disastrous for the next time something like this happens.

2. I'm extremely concerned about the effect the media has been able to exert on the general psyche. Culture has always been able to generate collective affect, but it's becoming worryingly uniform the more social media grows (I won't derail the thread by going into it here, but imagine starlings murmuring - one or two and it's just a mess going every which way, thousands and it suddenly looks like a choreographed dance, but all it is is just thousands of birds all trying to respond to each other and making tiny errors in copying which then get magnified)

It would be political suicide now for any government to act in a way which contradicted the media view (because it is so uniform) and any government which did want to lead (they're supposed to represent the population - not blindly follow it) simply don't have the means to spread information in the same way. It's not about political ideology anymore, it's about market-ready groups who can have focussed advertising delivered to them. Ideology has been subsumed into these groupings.

2,195 children every day die from Diarrhoea, 88% of which is avoidable by supplying clean drinking water and washing facilities. A relatively cheap intervention which doesn't even impact on issues like economic independence as other development aid might. The money to solve that problem is easily available, ready to hand and it really should have been sorted decades ago. any rational assessment of spending priorities would have focussed on it. But we don't get rational assessments of spending priorities when we jump from one media-instigated panic to the next.

When a plan is put into place and works, those opposed to it can always turn around and say it wouldn’t have mattered if no plan was used. — I like sushi

This is a non-sequitur. When a plan is put into place and the threat it was intended to avoid does not materialise we can say it was the plan, or we can say the plan was not needed. Neither is true or false out of the box. It depends entirely on the posterior analysis.

Yes. It depends entirely on the type and effect of comorbidities. The fact is that the overlap is unknown and will remain unknown until the end of the year (possibly even the following year, which I think is what Professor Spiegelhalter is referring to there).

My argument with @boethius is mainly about his ridiculous assertion that the overlap will definitely be small because there's no significant overlap in factors. This despite the fact the the only recorded factors affecting prognosis thus far are exactly the same as the factors affecting prognosis in other conditions, as the four articles I cited demonstrate.

I should add we already know a considerable amount about the effect of comorbidities from the death certificates. 91% in the UK and 98% in Italy. It should be stressed here as I think this has lead to some confusion these are not figures for "other things the patient had" which seems to be the prevailing opinion here. When we say comorbidity in this context we're not saying "Oh and he happend to have heart disease also, but that's irrelevant".

As the ONS specify "we analyse deaths involving COVID-19 by the main pre-existing condition. This is defined as the one pre-existing condition that is, on average, mostly likely to be the underlying cause of death for a person of that age and sex had they not died from COVID-19."

We're not talking about "and they also had..." we're talking about a condition that actually listed on the death certificate as a contributory cause of death and I stress - mostly likely to be the underlying cause of death for a person of that age and sex had they not died from COVID-19

I just don't know what more I can say to get this across.

It never ceases to amaze me the lengths people will go to to maintain their chosen narrative.

Apparently when Spiegelhalter uses the words 'many' and 'substantial', he means 'hardly any'.

Professor Ferguson, despite giving a cogent speech, was suddenly overtaken by an hallucination when he mentions a two thirds overlap.

Professor Spiegelhalter (a professor of mathematical statistics) apparently doesn't understand mathematical statistics.

Him saying the overlap 'is not the point' of the graph has somehow become him saying that there is no substantial overlap (oh, sorry I forgot 'substantial' now means 'very small' - I will have to get the hang of this newspeak)

I should never have started trying to have a reasonable discussion again.

Why don't you just take it up with the experts, they both have blogs. I can't be bothered with this condescending "I'll teach you where you've gone wrong" crap.

Prof Sir David Spiegelhalter {Professor of Mathematical Statistics at Cambridge), - "there will be a substantial overlap, Many people who die of Covid [the disease caused by coronavirus] would have died anyway within a short period," - he can be argued with at https://mobile.twitter.com/d_spiegel

Prof Neil Ferguson, the lead modeller at Imperial College London, has suggested it [the deaths of those who would have died anyway] "might be as much as half or two thirds of the deaths we see, because these are people at the end of their lives or who have underlying conditions.". - he can be contacted at https://www.imperial.ac.uk/people/neil.ferguson

When both Professer Spiegelhalter and Professer Ferguson have agreed that they're wrong and you're right, perhaps you'll have the authority to 'teach' the rest of us where we've gone wrong.