On Brazil’s strategy would the UK be 10x worse?

As Covid death rates in 2nd world countries (Brazil, India, Mexico, Peru, Argentina, etc) start to find their way into the militant contrarian narrative, like this shocking piece by Ross Clark, it become clear to me that some real data science needs to be applied to understanding their statistics. On the surface with deaths/1m rates of 638 for Brazil and 614 for the UK it looks like we’ve faired pretty similarly, with Peru looking worse at 946, but it is such a naive read to stop there and assume equivalence.

The population demographics are so different in really crucial way. Brazil has 3/10s of the proportion of people 80+, and we’ve seen 2/3 of our deaths 80+.

The UK’s median age is 40.5, Brazil’s 33.5 (India 26.8, Mexico 29.2, Peru 31, Argentina 31.5, all lower), and for a virus that kills 10+% of those 90+, and fewerand fewe the yougner you get, that Brazil’s is 8yrs lower (the others even more) has a huge effect on their average IFR. Sweden’s is 41.1, much more similar to ours, all 1st world countries are, which is why naive jurnos have not stumbled as much till now.

Surely IFRs are similar everywhere?

Yes, more so in a narrow band, say 45–49, we’ll find a much narrower range globally, healthcare access and socio-economic factors will make each band vary, but far less than simply being 85 not 45. However, no two countries have the same number of people in each band, and an IFR measured in one place includes an implicit weighting of the population it was sampled from. So use an IFR from the UK in Brazil and you’ll elevate the implied risk to their younger population. Do so in reverse and you’ll underestimate the risk to our more aging population.

It is the same innumeracy as blights Ioannidis’ work for the CDC. He just averages together IFRs from different studies. Each study will be a distinct blend ages in those sampled. If we globally tracked age banded means and stdevs, to understand the risk to a particular country we would just need to create a density weighted average to gauge the difference. This is the only study I’ve found that seems to be heading in this sensible direction.

So just how different are the populations?

For the UK, Sweden, Brazil and India here are the percentage proportions each 5yr age band makes up of the population in that country as a whole. What is immediately clear is just how much younger Brazil and India are than the UK or Sweden. In fact look at other 2nd/3rd world countries and you’ll see a similar distribution.

Is it just about the differences in age density?

No, everything is a factor, this is what they mean by heterogeneous. Consider two 45 year old men in both countries, are their chances the same?

1. You need to get it: the rate of spread will vary by age, and in each country each age band will differ due to lifestyle factors. Do old people life with famlies or in carehomes? What is the population density, transport system, etc. Favelas have population densities of 100k/sqkm, Islington the densest in the UK is 16.1k/sqkm, Stockholm only 4.5k — all these differences lead to an average r0. Ferguson’s model tries to factor all this in nationally, and as I discuss in this piece it is little know his March paper also included a 48k prediction for this scenario based on this.
2. Even if you get it: being a man not a women seems to make you 50% more likely to die. Add to this comorbidities by age likely being worse in Brazil, leading to lower life expectancy, means perhaps a 45yr old is more like a 50 yr old in the UK. That said obesity (a Covid risk factor) is more of a problem in the UK, so perhaps things cancel each other out.
3. Will you get help? In Brazil there is at least universal healthcare, in other 2nd/3rd countries the figures we see may only correspond to the 1/3 of those with health insurance, who may be a “far less likely to die” part of the population. So you just can’t line numbers up. At least this is easier with Brazil, but even so it is likely even more skewed to better provision in cities than the UK, and with only 2.2/1000 doctors vs the UK’s 2.8, supply is more constrained. Further, the healthcare outcomes are probably worse.

Reality is all these factors matter, but the difference between being 45 or 80 will be far bigger, so let’s make sure we start in the right ballpark.

So, what does the age density imply to each countries IFR?

To explore let’s use the same r0 (and ceiling of 65%) for all countries and age bands. Then, using average age banded IFRs (that I could find from various global studies averaged together), along with age banded population totals for each country — here are the estimates of the infections and deaths we would see in each 5yr band, summarised into coarser bands. From these we can then form country specific IFRs.

1. Brazil & India are dramatically lower than the UK, 2/5 and 3/10 each.
2. Sweden is slightly higher than the UK due to a slightly older population.

Bear in mind, these ratio is unaffected by the ceiling, set it to 20%, you’ll get the same ratios — but, what is really important the capacity for there to be that magnitude of difference between two countries.

So, Brazil’s 638 d/1m (if the spread is even) could be 1595 if they had the UKs demographic, which is 2.6x our 614.

Can we find data to see in more detail by age band?

The problem with this estimate is it assumes spread has followed similar patterns. So can we find age banded mortaility data to see the real difference?

Even in 1st world countries age banded data is hard to come by, but thanks to a defiant health service similar to the NHS, and a journalist called Marcelo Soares’ struggle to keep the data freely available in Brazil (despite Bolsenaro’s attempts to muzzle it, this article covers this struggle well) there is quite extensive and detailed data available for the whole of Brazil on this very excellent website.

From this I have compiled the plots above, they show:

1. Age density side by side: those 80+ make up 1.5% of the Brazilian population, 5% in the UK, 64% are below 40 only 50% in the UK.
2. Proportion of Covid deaths in each band: having a 3/10 the number 80+ means only 27% of Brazil’s deaths are 80+, in the UK 62%. They have 25% of deaths <60, the UK only 7%.
3. Ratio of deaths/1m (blue line): shows the ratio of the rate of death in each age band — e.g. 40–49, 296/86 = 4.67x. The pattern is the younger you get the higher the ratio. Most sobering <20 they’ve seen 16x the rate we have. Yes, they have 3x our population, and 30% <20, not our 23%, but in real terms they’ve seen 207 (0–9) and 861 (10–19), compared to our 4 and 12. 80+ the ratios are not as high, they done relatively better, but it still suggests our care home epidemic could have been much worse.
4. Ratio of the % of deaths (green line): as they have more people in younger bands, this shows the multiplicative effect on the share of deaths.
5. Ratio of pop. %: gives you some idea of the greater or lesser degree to which each age band contributes to the IFR calced above.

Now, deaths/1m are a good proxy for spread, if we use this age banded detail it suggests as per the final plot that Brazil has seen 7.1x our spread, as in far higher than the ratio of 2.6x in overall deaths, and this is entirely down to far higher rates in the young, but them being far smaller in absolute terms.

What does this tell us about the ceiling for infection?

Essentially don’t believe the hype that all policies here (for sure many were over applied) were pointless. Brazil suggested similar measures to its people, but has seen far less compliance than even Sweden. As I show in this research they have had 3x our mortality on a like for like basis.

Opposite to what Ross Clark concludes, I don’t see 3 countries with similar d/1m, I see 3 countries with very different spreads, that coincidentally average up to the same. If he looked at the wider picture he’d see most other countires are very different, he is falling for the very definition of confirmation bias, find a small number of examples that support your gut feeling. There is no 20% Tcell driven herd immunity ceiling that we were always going to hit regardless of government policy, it is far more complicated than such a naive narrative.

Brazil’s likely hidden deaths

So are we done? No, we haven’t really considered Brazil’s excess mortality? It is hard to get national data for most countries, and this offical source for Brazil still has nothing. Some cities have released their own. The Economist has gathered and shared this data on github leading to this report, which seems to suggest outside of really westernised cities like Rio, say Manaus, test counted cases may be as bad as 1/5 that of the excess deaths.

Here are Brazilian and UK 7 day rolling averages of daily new case and deaths normalised to per 1m for comparison.

On deaths, their numbers have plateaued since late May, we know from anecdotal reports hospitals hit peak capacity at this point, suggesting, unlike the UK, Brazil has been denying treatment, but without all mortality data we cannot be definitive as to how many, but as an inequality, it is likely to elevate mortality more than our excess figures do, so it will not paint a better picture.

On cases, their daily numbers hit 3x our peak, but they do little community testing or test&trace, so comparison is tricky. Plus, to get to that level they there have been through 3 plateaus, suggesting they have periodically hit limits, then added on more streams of testing, so probably more than we have been, capacity constrained in just testing the most sick. It looks like they have peaked, but they have months to burn out.

Essentially, our excess mortality figures elevate our deaths from 50%, it is entirely likely given all these greater deficiencies in recording, that they could easily see excess deaths at least double recorded rates, possibly triple.

Do we need a final worry about density?

Now, as I raise about Sweden in this article, density in this comparison matters, and São Paulo is denser than London, and with nationally 7% of the population living in Favelas which are 5x denser than our densest places, their spread in these areas always going to be higher. So perhaps we need to temper this ratio down?

For sure I’d love to investigate this more, but my observation would be outside the centre of a handful of cities they’re consistently less dense than the UK. So what is their average natural degree for spread relative to us? Perhaps the simplest indicator is their infection has grown far more slowly than it did in the the UK, suggesting their average natural unpsressed r0 was lower, so if anything this estimate would need adjusting up again to account for this.

What is the real takeaway here?

Direct comparisons are naive as Brazil is not measuring consistently, and does not have the same age demographic. At the moment they are at 104% of our d/1m, but to get to a comparable estimate we need to:

• x1.5: to get them from their peak to a burn out state
• x2.5: once all mortality figures are published
• x2.5: to make the numbers demographically comparable to the UK

… and you quickly get to Brazil looking like 9.3x as many people, on a like for like basis, having died to burn the virus out, below 40 this could be well above 10x by the end.

I’m not trying to argue in support of government policy. Simply trying to remove false argument from debates. I think nightclubs should be open and the people of Oldham should be made aware of a need for caution as cases are elevated, but being able to choose to go to a nightclub and avoid their parents, rather than not having the choice at all. However, I believe you win no argument to restore these freedoms by demonstrating we have no grasp of the reality of the data. The UK has 41k tested deaths and 61k excess deaths. On Brazil’s path it is looking like this would have been more like 350–550k deaths… sound familiar?