One flu over the cuckoo's nest

And this bird you cannot change

There’s a nice explainer of the current bird flu situation in today’s Times by Tom Whipple. There’s a good account of the dangers of raw milk, the situation with vaccines, and the potential for future dangerous mutations. However, there’s also talk of a 50 per cent fatality rate, which is a number that is doing the rounds.

Whipple rightly raises some scepticism about this:

Does it really kill 50 per cent of people?

We don’t know. When people at a poultry farm get mildly ill, doctors rarely conduct a full genomic analysis of their cold. It is only if they are seriously ill that we take notice, and realise it’s something different. Only then are they logged as a bird flu case. So it’s possible that some, perhaps many, cases are being missed — and that by only picking up the worst, we exaggerate the worst effects.

But I think his explanation is fine, but undersells it - I’d say we probably do know it won’t be 50 per cent. And that’s a massive relief, because of course a 50 per cent death rate would be incredibly bad. That’s the kind of number quoted for Ebola by the WHO:

The average Ebola case fatality rate is around 50%. Case fatality rates have varied from 25–90% in past outbreaks, depending on circumstances and the response.

However, because Ebola is so dramatic, involving total collapse of bodily organs, and because it’s taken seriously by doctors, outbreaks tend to be relatively contained. For example, this paper quotes a total of just over 15,000 Ebola deaths between 1976 and 2020.

While this is obviously bad, and should not be ignored just because the deaths don’t take place in richer countries, it’s nothing like the effect we’d see from a virus which spread widely like COVID and had the same sort of fatality rate. There, without vaccines and therapeutics, there’d be tens of millions of deaths in the UK alone. Indeed, while some of the arguments from the lockdown sceptic side were based on the idea that COVID wasn’t a “proper pandemic”, because a pre-vaccine fatality rate of the order of 1% wasn’t worth worrying about in their view, that still converted into hundreds of thousands of deaths as we know.

So the idea of a 50 per cent fatality rate in humans for bird flu should certainly make us sit up and take notice. And indeed, at a glance, this doesn’t seem so far wrong: the WHO have been collecting data over the last 20 years, and you can see here that 888 reported cases have resulted in 463 deaths, or a case fatality rate of 52%.

However, as Whipple suggests, there’s a problem with this. As in the COVID pandemic, we have to remember the difference between case fatality rate and infection fatality rate. The former will be higher, because not all infections will result in a positive test (a “case”). Indeed, if a relatively small fraction of infections result in a positive test then this can dramatically overplay the risk.

Take COVID for example. If I go to the UKHSA dashboard then I will see that the latest figures are 2,343 cases and 92 deaths in a week. Crudely dividing one by another suggests a case fatality rate of about 5%. Doing a better estimate, taking into account lags and rising and falling trajectories, the most recent deaths probably relate to a time around early April when cases were perhaps half that. So we could perhaps even argue for a case fatality rate in the 8-10% region.

And yet, I already mentioned that pre-vaccine the COVID infection fatality rate was much lower than this (perhaps of the order of 1%). Now, it’s likely that the true value is perhaps somewhere in the 0.01-0.05% range. In other words, our dashboard-derived case fatality rate estimate is something like 200 to 500 times too high.

There’s a simple explanation for this: nobody is testing any more! The vast majority of reported cases arise at the hospital admissions stage, so are hugely skewed to the most serious infections and the most vulnerable people. People for whom COVID is somewhat like a case of flu don’t tend to show up in the data - even those who do take an lateral flow test don’t have an easy mechanism for reporting the result.¹

So, returning to bird flu, I believe that it’s likely that similar things are going on (as Whipple suggested). Looking again at the WHO table you can see that the vast majority of the reported cases (861 out of 888) took place before 2014, the majority of them in countries like Indonesia, Vietnam and Egypt which locally reported a very high case fatality rate at the time. While this data is valid in a sense, it’s not clear to me that an estimate which is heavily weighted to decade-old estimates of a healthcare-dependent quantity (dominated by these kinds of countries) is representative of what the experience might be in the UK now.

Indeed, Whipple suggests that the estimate might be lower if we did surveillance of all farm workers, rather than just picking up the most serious cases. But that has been happening! For example, the most recent UKHSA surveillance report describes 4 cases picked up by random sampling: 3 did not even show symptoms, and 1 reported a sore throat and myalgia. While four people is a small sample, and so random chance will play a role, it seems hard to reconcile these numbers with a true fatality rate of the claimed 50 per cent.

Of course, even if the fatality rate were much less than 50 per cent, that doesn’t mean that a widespread avian flu pandemic wouldn’t be bad. If the figure was 1 per cent, that could still be extremely serious if the outbreak became widely established, not least because we wouldn’t see anything like the age skew of COVID. As I mentioned here this rate of deaths among children would make for some extremely difficult conversations around school closures and other measures.

So it’s definitely right that we take bird flu seriously, and that infectious disease experts should be planning ahead to mitigate the risks. But equally I don’t think it helps anyone to be routinely quoting a science fiction sounding fatality rate, without at least thinking a little bit about what that means and how it is derived.

1. You can also see Our World In Data for a nice explainer of all this.