Thanks for the suggestion. I’ve put it up as a post now.
Radford Neal
Karma: 736
I certainly agree that BMI is not ideal, but that’s what we’ve got as far as most presently-available studies go.
In addition to all the problems that you mention with BMI, there is just systematic measurement error, which is seldom mentioned. Is the height measured with or without shoes? Do unhealthy people tend to slump rather than stand up straight, reducing their measured height, or does the person measuring height make sure to get them to stand up straight? Is weight with or without clothing? What time of day is weight measured (the morning, right after a big meal,...)? I think different answers to these questions could shift measured BMI by about 0.5, perhaps systematically for different studies. Of course, this would affect ABSI too, along with whatever similar problems there are for measuring waist circumference.
Thanks for the link to the paper on ABSI. It looks very interesting. From a first glance, it seems to support the view that the WHO “normal” category for BMI extends too far in the low direction. (Of course, using ABSI as well rather than just BMI might give better information.)
If one has measurements on height, weight, and waist circumference, and a reasonable number of subjects, I think one would actually want to look at all three measurements, not just a one or two dimensional condensation of them, using some suitably flexible model for the relationship with mortality.
Yes, an R0 (or maybe Rt is the term to use) of one would give linear growth. But why should it be close to one in so many countries? (I think 0.85 isn’t close enough.) There seems to be no reason for the effect of the interventions that have been made to hit one this closely, other than shear coincidence.
One of the arguments I’ve heard against “flattening the curve” is that to keep infections below health care capacity you have to get Rt so close to one that you might as well do a bit more and get it well below one. (And anyway you’d have to aim for that to be sure that it doesn’t stay well above one.) It’s hard to believe that we’d hit one so precisely when nobody knows what the effect of the interventions really is.
I discuss that possibility in the post. It seems more plausible as an explanation for linear growth in confirmed COVID-19 cases than for linear growth in deaths attributed to COVID-19. For the latter, the situation would have to be something like… there’s no tendency to use tests on the most serious cases (many of whom die), and instead they test a fixed number chosen at random of those who come to medical attention (of whom the most serious are a small, fixed fraction), and only attribute a death to COVID-19 if it was of someone who tested positive.
Thanks for the interesting graphs!
Rt=0.85 with serial interval of 6-8 does look almost like a straight line for the relevant time period. Given that the actual data is noisy (probably beyond simple Poisson variation, with various reporting effects), it may be compatible with that explanation (without, for instance, needing to hypothesize stranger reporting artifacts that would systematically keep the reported deaths nearly constant). Though the linear plots of world case and death counts at https://www.worldometers.info/coronavirus/ do still look very straight to me.
As a more general point, it’s not entirely satisfactory to say that you made an observation and got Rt approximately one, so that’s just what it is. The simple model would be that initially R0 was something greater than one (otherwise we’d never have heard of this virus) as a result of viral characteristics, human behaviour, weather, etc. - it could be 1.3, could be 4.7, etc. - and then we changed our behaviour, and so Rt became something smaller than R0 - maybe a lot smaller, maybe a little smaller, hard to tell. There’s no reason in this model that it should end up really close to one, except by chance. If it seems to be really close to one, then alternative models become more plausible—such as a model in which testing or hospital limits somehow lead to reported cases or deaths saturating at some upper limit (regardless of the real numbers), or in which the transmission mechanism is something completely different from what we think—since in these models there may be a good reason why the apparent Rt should be close to one.
I could see that as the explanation if the curve has stopped going up exponentially because we are approaching herd immunity, but that isn’t what’s commonly believed… Instead, we’d expect the number of new deaths each day to have shifted to a different exponential curve as a result of control measures, and only by coincidence would this exponential have a coefficient near zero (hence nearly constant deaths).
[Though as Vaniver points out, it’s a bit more complicated when you account for the serial interval (the time lag for an infected person to infect another).]
I’m not following the logic here. I presume that “fast takeoff” is supposed to mean that someone with increased intelligence from the first improvement is then able to think of a second improvement that would have been beyond what earlier people could have thought of, and so forth for additional improvements. The relevant time interval here is from birth to thinking better than the previous generation, which need have nothing to do with the interval from birth to reproductive maturity (an interval which is not immutable anyway). The person who thinks of the new improvement doesn’t have to be one of those who gestate the next generation.
Those are both good points.
It does seem likely that the measures that have been taken do a lot more to inhibit between-household transmission than within-household transmission, so there could be a lag while within-household transmission works itself out before an exponential decline in new cases/deaths becomes evident.
And if there’s a testing backlog, and deaths are only recorded as COVID-19 deaths once a sample is finally tested, that would also introduce a lag. This seems like it must vary a lot with locale, though...
No, I wasn’t thinking of modification of adult somatic genes. I was thinking of reproductive maturity taking 12 years, which you’re right is also about how long it takes to reach adult levels of cognition (though not knowledge, obviously). The coincidence here leads to the ambiguity in what you said. Actually, I doubt this is a coincidence—it makes biological sense for these two to go together. Neither would be immutable if you’re making profound changes to the genome, although if anything, it might be necessary to prolong the period of immaturity in order to get higher intelligence.
I looked at covid and BMI in a post linked here. Using data from UK critical care units, it seems that BMI from 30-40 is only a small risk, though the risk from BMI>40 is substantial. This is after adjusting for age and sex.
A more recent UK study here found increased risk of death, after adjusting for age and sex, for obesity (BMI>30), and a substantially higher risk for BMI>40. These risks went down a bit when also adjusting for other factors. (Though depending on your purpose, this may not be valid, since some of these factors might be on the causal path from obesity to risk of death.)
You’re saying that obesity matters mostly if you’re younger than 60, which is a question not directly addressed by these results (which come up with a single risk ratio, on the assumption that it’s the same for all ages). The study here which you link to gives a scatterplot of BMI versus age in covid patients, showing a negative correlation. This is not meaningful information without a corresponding scatterplot for the general population from which these patients came, which they oddly do not show (or even speculate about, if data is hard to find). In that scatterplot, patients with really extreme obesity (BMI>50) seem to play a large role.
In my post, I present evidence that low BMI may be a risk factor, perhaps not only for BMI below the official “underweight” threshold of BMI<18.5 but also for around BMI<20. Frustratingly, many studies do not present the data that would allow one to investigate this. For the recent UK study I linked above, they use “not obese” as their reference class, grouping everyone with BMI<30 together, though their raw, unadjusted, data indicates that there is a substantial firsk for BMI<18.5 (and even in that data, they group everyone with BMI from 18.5 to 24.9 together, even though past studies show that there may be substantial differences within this supposedly “normal” range).
When looking at varying effects by age, one possibility to consider is that high BMI is a risk factor for younger people, and that low BMI is a risk factor for older people (as opposed to BMI not mattering for older people, as you sort of conclude).
That seems like a plausible explanation. I wonder how one could confirm it? If there were two sets of covid death counts based on strict and on loose criteria, one could see if there is a divergence in their rates of increase.
Your comment didn’t show up at my blog. Not sure why...
From the abstract: “The incidence of new illness compatible with Covid-19 did not differ significantly between participants receiving hydroxychloroquine (49 of 414 [11.8%]) and those receiving placebo (58 of 407 [14.3%])”
So the treatment group did have a lower incidence of illness than the control group, but the difference wasn’t statistically significant. However, only 107 patients in total became ill. This is a rather small sample, so the results by no means rule out a clinically important benefit of HCQ. Even just taking the observed proportions as a best estimate, there’s a 17% reduction of illness in the treatment group, which doesn’t seem negligible, and the actual benefit could plausibly be considerably larger. (Of course, given the small sample size, it’s also plausible that the real effect is in the other direction.)
Indeed. And one can come up with other lists that stem from somewhat different moral intuitions, like:
Forcing children (even against the wishes of their parents) to spend much of their day being indoctrinated in whatever ideology is dominant in university departments of education.
Confining prisoners (many not guilty of anything that should be a crime) in conditions where it is guaranteed that they will be victims of multiple criminal acts.
Confiscating about half the wealth produced by people, for the benefit of politically favoured insiders, or to be distributed to subsets of the population in order to buy political support.
Subverting freedom of speech by heavily regulating use of first the radio spectrum, and then when technology changed, wired electronic communications, continuing with this even when such media become the dominant form of communication.
Preventing sick people from taking whatever drugs they think have the best chance of saving their life.
So one can see that there are plenty of things currently supported by large numbers of people that are plausibly in the “worse than Hitler” category, without even getting into possible future denigration of the colour green.
And of course the opposite of all the above might also be plausibly condemned by some future society:
Allowing children to be misled by their parents’ incorrect views.
Letting too many bad people run free when they ought to be in prison, and not some wishy-washy “nice” prison, of course.
Letting rich (>$40000/year) people spend their money on whatever stupid thing they want, rather than on things that are socially useful.
Letting people express harmful opinions.
Letting sick people take drugs that don’t work, or if they do “work”, do so only by saving their life but leaving them weakened, and a burden on society.
There’s really no substitute for making your own moral judgements. The idea that the future will always be more moral than the past seems quite false. Even if there is some slow, long-term moral progress (which I think may be the case), there are obviously significant regressions over the time scale of decades and centuries. Going by what you think (correctly) will be the moral views 20 years in the future would not be a good thing in the Germany of 1920.
Another possibility is better disease survival due to increased vitamin D levels from sunshine (or due to some other physiological effect of sunlight).
The effect seems rather large for this to be the explanation, but it sure would be great if a bit of sunlight is all that’s needed!
Thanks! Very interesting.
It would be great if the mouse results turn out to apply to humans as well, but I have my doubts. These doubts are based on what I thought were pretty conventional biological assumptions, but that nevertheless don’t seem to be addressed in the anti-aging discussions I’ve seen.
The basic problem is that there’s a good reason mice don’t live long. Even if they didn’t age, the environment in which they live means they are very likely to die in a few years from starvation or predation. So genes that keep them from aging won’t be selected for because of either or both of two reasons: (1) The selective advantage of not aging, when you’re likely to die young anyway, isn’t enough to overcome random mutation that undoes the anti-aging genes. (2) The advantage of not aging comes at some (possibly rather small) cost in terms of increased likelihood of death from predation or starvation, or decreased fecundity early in life. (For instance, it might have an energy/food cost, or might come with decreased physical performance, such as in running speed.)
Humans live in a different environment, in which slower aging is more advantageous. And indeed humans age much slower than mice, presumably because we have genes that enable various anti-aging strategies that mice lack.
So, when a drug is found to slow aging in mice, the first question in my mind would be, “is this drug enabling a mechanism that is already present in humans?”.
And the default answer to this question would seem to be “yes”. If there’s some simple biochemical way of slowing aging, why don’t humans already have this, given that slower aging in humans would give a significant selective advantage? (Even (especially?) in pre-civilizational societies, significant numbers of people die of old age rather than from violence or starvation.)
On this reasoning, one would expect that a successful anti-aging program would have to involve something complicated, not easily produced by evolution. Something like, for example, nanobots inspecting cells for damaged DNA (comparing against a consensus sequence derived from a large number of the person’s cells), and killing cells that are too damaged. Or at least, if there is some relatively simple intervention that helps, one would expect it to be sufficiently subtle that it doesn’t show up in mice (but only after decades of life, when selective pressure for it in humans is comparatively small).
I think some of these examples are not real examples of “inadequate equilibria”. They are instead situations with real switching costs, or where there are conflicting beliefs or interests.
To illustrate, the authors example of moving from proprietary to open-source journals seems like a real example to me. But their example of using Bayesian rather than frequentist statistical methods does not seem like a real example.
Note that I’m a Bayesian (though not a rabid one), and that I’ve taught introductory statistics. There isn’t some easy way to just switch to Bayesian methods. First, students need to understand the scientific literature, and that includes the past scientific literature. So for a considerable period of time, students will need to understand frequentist statistics. This is a legacy compatibility problem, not a coordination problem. Second, there are many scientists who don’t know Bayesian statistics. This is a retraining problem (which is not cost-free), not a coordination problem. Third, not everyone agrees that Bayesian methods are better. This is a persuasion problem, not a coordination problem. Fourth, Bayesian methods aren’t always better—there really are problems where the correct Bayesian approach is much, much more difficult to carry out than a simple frequentist approach that is usually gives mostly-correct results. So it really is necessary for at least some people to understand frequentist statistics, though it would be good if the emphasis eventually changes in a Bayesian direction. There may be some coordination aspects to the current mess, but to a considerable extent the mess reflects real issues with real costs and benefits.
It’s not clear what happened at Dodger Stadium. You say that “Far right anti-vaccine protesters blocked access to the mass vaccination site at Dodger Stadium, forcing it to shut down for a day.”. But if you read this account:
and try to read between the lines, it’s not clear. They say that “The Los Angeles Fire Department shut the entrance to the vaccination center at Dodger Stadium about 2 p.m. as a precaution”. That doesn’t actually sound like the protesters prevented people from entering. It sounds more like they waved signs and shouted, and the authorities thought that they might become violent, so they shut things down (for an hour, not the whole day).
The news report you link to on the CDC study of BMI and Covid is highly misleading. The CDC report (https://www.cdc.gov/mmwr/volumes/70/wr/mm7010e4.htm) says
“the analyses in this report describe a J-shaped association between BMI and severe COVID-19, with the lowest risk at BMIs near the threshold between healthy weight and overweight in most instances”.
The CNBC news report falsely states that “The agency found the risk for hospitalizations, ICU admissions and deaths was lowest among individuals with BMIs under 25”. They go on to say, “It doesn’t take a lot of extra pounds to be considered overweight or obese. A 5-foot-10-inch man at 175 pounds and 5-foot-4-inch woman at 146 pounds would both be considered overweight with BMIs of just over 25″, without mentioning that this BMI of 25 is actually optimal for avoiding hospitalization for all but the youngest (18-39) age group, according to the CDC study (see Figure 2).
I’ve written a blog post on “Body Mass and Risk from COVID-19 and Influenza”, available at https://radfordneal.wordpress.com/2020/04/06/body-mass-and-risk-from-covid-19-and-influenza/
Here’s the intro:
Understanding the factors affecting whether someone infected with COVID-19 will become seriously ill is important for treatment of patients, for forecasting and planning, and — with factors that can be changed — for personal decisions aimed at reducing risk. Despite our current focus, influenza also remains a serious disease, so understanding its risk factors is also important.
Here, I’ll look at some of the evidence on how body mass — formalized as Body Mass Index (BMI, weight in kilograms divided by squared height in metres) — influences prognosis for respiratory diseases. Information specific to COVID-19 is still scant, but there is more data on influenza and on other respiratory infections (which includes coronaviruses other than COVID-19). Information on how BMI relates to general mortality should also be helpful.
Below, I’ll look at two relevant papers, plus a preliminary report on COVID-19. To preview my conclusions, it seems that being underweight and being seriously obese are both risk factors for serious respiratory illness. Furthermore, it seems that “underweight” should include the lower part of the “normal weight” category as defined by the WHO. Official advice in this respect seems dangerously misleading.