Posts written by Thomas Lumley (2534)

Circulating on Twitter, but originally from US News and World Report

It’s an interesting list.

Most of the people talking about it on Twitter wanted to ridicule the list without actually worrying about how it was constructed, so it didn’t come with any link or any explanation beyond the source. It’s not hard to find some information, although I haven’t been able to get full details.

There are different ways you might go about constructing a ‘racism’ ordering for countries. According to a 2013 story in the Washington Post, one ranking of basically this sort started with a question in the World Values Survey. Two researchers (I’ll let your prejudices work by saying they were Swedish economists) wanted to look at relationships between economic freedom and racism.  They needed something widely measured, and used a question about kinds of people you would not be happy with having as neighbours.  One of the options was “people of a different race”, others include “people with AIDS”, “immigrants”, “heavy drinkers”, “unmarried couple living together”, “people of a different religion” and so on.  These economists used as their metric for racism the proportion of people who would not want someone from another race as a neighbour.  If you were being pedantic, like me, you might call this a xenophobia/xenophilia score rather than a racism score. It clearly measures something relevant, but you’d expect it to miss the “Some of my best friends are black/gay/Jewish/etc” type of polite racism. This follow-up piece at the Washington Post  covers some of the other complications.

The scale based on the World Values Survey has some agreement with the current version, but it’s not the same. In particular, the USA does quite well on the World Values Survey question, but rates low on the current metric.

The current version is from a survey called “Best Countries“. It has a simpler structure. Respondents (10,068 were informed elites, 4,919 were business decision-makers and 5,817 were considered general public) rated each country on 76 attributes, one of which  was racial equity.  They were also asked whether they agreed “A country is stronger when it is more racially and ethnically diverse” but it doesn’t appear this goes into the ranking (the Danes and Swedes were below the global average on this question, though NZ and Canada were high).

So, the ranking is based on whether a sample of people around the world, targeting ‘informed elites and business decision makers’, thinks that the country is racist or not. The problem with a ranking like that is that most respondents have no actual idea of whether Denmark or Botswana or Agrabah or Paraguay is racist; they’re just going by their own prejudices and what they see in the news.  It’s quite likely that the very low rating for the US is due in part to the Black Lives Matter protests — which you could argue were a good sign not a bad sign for US attitudes on race.

From the NZ Herald, squashed-trees edition

It’s not really clear what’s going on here: the 3.75% at the bottom right vs the 3.75% at the top left.

Things are better on the NZ$ Herald website, under the headline The great divide: Why are NZ interest rates so much higher than Australia?

Here it’s clear that the label at the bottom right had just gone feral somehow and that the graph is at least plausibly correct. There’s still a bit of a problem in that, at least for the historic part of the graph, the lines should be flat where they don’t jump; there shouldn’t be any slopes. The RBNZ didn’t come out in early 2010 and say “we’re going to smoothly decrease the rate from 3 to 2.75 over the rest of the year”; that’s not how they work. Also, NZ interest rates aren’t actually “so much higher” than rates across the Tasman; they’re just projected to be higher.

Checking against this July graph from interest.co.nz basically confirms the numbers, though there is some interesting disagreement if you care about details, such as the shape of the interest rate rise and fall in 2014-15 and whether the Oz rate was above or below the NZ rate at the start of 2016

The July projections diverge less than the current predictions do: the banks aren’t actually all that good at predicting interest rates two years ahead.

The spurious slopes are still there in the graph, though in this one at least the flat bits are flat and it’s just the vertical bits that aren’t vertical. That’s even a problem on the official RBNZ website.

None of this is a criticism of the actual content of the Herald piece, which both talks about the reasons for divergence and quotes experts who don’t think the diverging forecasts will hold up

Ultimately, despite the two very divergent central bank views, the answer is somewhere in the middle and the rate tracks will move closer in the year ahead, McLeish says.

But the graph and headline don’t help

From Sophie Jones on Twitter:

If you zoom in on the fine print, that’s 50.6% of 15096 people preferring Pepsi Max over full-sugar Coca-Cola.  You could quibble about the comparison — should this be restricted to cola drinkers (or non-cola drinkers); what happened to the diet versions; how about L&P? — but it’s a comparison.

More obviously, 50.6% is very close to 50%.  You  might ask what the margin of error was for a sample of 15000. It’s more than 0.6%: these results are consistent with just a coin toss.  It might taste like victory, but only if victory doesn’t taste very distinctive.

On the other hand, Pepsi lost the cola wars in New Zealand, so the starting point might reasonably not be 50:50.  This survey doesn’t convincingly show that Pepsi Max is preferred over Coca-Cola by a majority even in blind two-way comparisons, but it does show it’s not far behind. And, in context, that’s probably worth advertising.

The Covid vaccine is safe and effective and it’s good that most eligible people are getting it. But how much protection does it give? If you look at the NZ statistics on who gets Covid, it seems to be extraordinarily effective: the chance of ending up with (diagnosed) Covid for an unvaccinated person is about 20 times higher than for a vaccinated person.

That’s probably an overestimate. People who are vaccinated are at lower risk for immunological reasons: the vaccine really works.  We’re also at lower risk for social reasons: if you’re vaccinated, your friends and family and people you interact with are also more likely to be vaccinated, so they are less likely to give you the virus. That’s partly due to equity problems in the vaccine rollout and partly just to what social-network people call homophily:  you tend to hang out with people similar to you. The immunological reason will hold true over summer; the social reason perhaps less so if people travel. 

Also, because elimination came so close to working in Auckland, the virus has been fairly effectively suppressed in most of the New Zealand population.  On top of the clustering of unvaccinated people, there’s very strong clustering of the current outbreak — it’s mostly in Auckland, but it’s not at all evenly spread within Auckland.  Even if you’re in Auckland you probably know either no-one or lots of people who have been infected.  If you’re in know no-one, you’re at lower risk– and you’re probably vaccinated. As we go from a more or less localised outbreak to many little outbreaks, this additional clustering will go away and the apparent benefit of vaccination will fall.

How much will it fall (and why am I sure)? In the USA, you’re currently about 6 times as likely to get a Covid diagnosis if you’re unvaccinated (according to the CDC). In the UK, the ratio comparing unvaccinated people to those with a Pfizer vaccination within four three months is 4-5.  That fits with the estimates of how effective the vaccine is, biologically, against Delta, plus a bit of social clustering.  The ratio in NZ will be heading that way over time.

So: vaccines, yes, but also masks and distancing and meeting people outside when you can and getting tested if you have symptoms and not going to isolated places that don’t even have enough of their own health care.  Don’t give the virus an inch.

The external scientific advisory committee for the FDA meets next week to consider the Pfizer Covid vaccine in kids 5-12.  Pfizer’s briefing to the committee is now up on the website; the FDA briefing is not (as of midday Saturday).

Demographically, the trial isn’t as representative as the initial adult trials, which were much better than usual. About 10% of the participants had asthma and about 10% were obese. Black and Hispanic populations were under-represented by about a half relative to the US population —  though there has been no indication that race/ethnicity matters for vaccine efficacy so far.

For the extension to ages 5-12 there are basically three questions

1. Is the dose right? They used 1/3 of the adult dose. Using too much would increase  adverse effects; using too little would not provide reliable immunity
2. Is there anything new and worrying about adverse effects? This trial, like all randomised trials, is too small to see surprising and rare adverse reactions that happen to 1 person in 10,000 or 1 in a million. These can only ever be picked up by post-marketing surveillance, as we’ve seen for both the Pfizer and AstraZeneca vaccines. Safety signals in the trial would involve milder, less rare reactions at elevated rates.
3. What is the risk/benefit relationship like, given the relatively lower risk from Covid in kids and the fact that (in contrast to many infectious diseases) kids don’t seem to spread it more effectively than adults do?

The data are positive on the first point. Pain and redness at the injection site are a bit more common than with teenagers given the full dose; systemic reactions such as fatigue and fever are a bit less common.  Levels of neutralising antibodies are about the same as in teenagers given the full dose.

On the second point, the trial is again positive. Nothing new seems to have been seen (though this is where the FDA briefing will be important, to see if they agree — how you classify adverse events can make a difference).

It’s harder to say what regulators will think for the third point, but if the FDA were willing in principle to approve based on a trial of this general design there doesn’t seem to be anything obvious in the results that would make them not approve based on these data.

Update: FDA’s briefing is now available.  The main new information is an explicit risk-benefit analysis. As you’d expect, the net benefit depends on the Covid incidence, but they say the net benefit might be positive even under the lowest-incidence case (and that’s assuming only 80% effectiveness against hospitalisation, which is probably a bit low, and a pessimistic view of the data on myocarditis). They don’t seem to model any community effect of vaccination by reducing infection in other people and thus reducing exposure to the virus. Risk-benefit is where the most interesting discussion should be at next week’s meeting.

It’s time for NZ’s Bird of the Year.  There are two important things to remember about Bird of the Year. First, if you’re going to talk about it on social media use the appropriate hashtag so normal people can mute you.  Second, Bird of the Year is a popularity contest driven by who votes and by last minute social influences; it doesn’t tell us anything new about the birds themselves or really even anything about their popularity in other settings. That’s not the thing you need to remember. The thing you need to remember is that all other online bogus clicky polls work the same way.

I’m campaigning for kākāpō because Zoe Luo, who is doing a PhD with me, is working on ways to model the genetics of rare species, using kākāpō as an example.  The entire kākāpō species had full genome sequencing done.  You usually can’t do that; Zoe is looking at how to use genome sequencing on a sample of the population together with other information on the rest of the birds to fit similar statistical models to the ones you would fit with full genome sequencing.  We can do this with the kākāpō because it’s easy to see what how well your models would work if you had less data than you really do — you can just ignore some of it — but it’s harder to see how well your models would work if you had more data than you really do.

Today is the NZ Vaxathon! I had my second dose last week. All the statisticians I’ve talked to on the issue have had either one or two doses depending on when they became eligible.  For fans of anecdata, our experience varied from no adverse effects at all to a day or so feeling pretty bloody average, consistent with what the randomised trial reported.

In the spirit of StatsChat, here are some links to primary sources for anyone who wants to do their own reading rather than getting clinicians and scientists to translate it

• Official guidance from the US Food and Drug Administration on what sort of evidence they would need to approve a Covid vaccine (June 2020)
• The protocol (plan) for the Pfizer vaccine trial. It’s unusual for these to be public, but it was reassuring when this one was published last year (PDF)
• The meeting briefings for the external expert advisory committee meeting that recommended authorisation of the vaccine. You want ‘Briefing Document -FDA’ and ‘Briefing Document -Sponsor’. And if you’re really dedicated, the transcript of the discussion.  This is where you get all the detailed information that doesn’t fit into a published research paper, plus discussion from the medical and scientific experts
• Real-world effectiveness assessment of the vaccine in Israel. Authors include Miguel Hernán, who is a leading expert on causal inference from observational data
• What we know about heart inflammation after vaccination: an adverse reaction in about one person per 100,000 aged under 40.  Most recover fully, but there has been one death in New Zealand that was probably due to this.
• A summary on vaccine effectiveness with links to primary sources, covering effectiveness against Delta, prevention of infection, and prevention of symptomatic disease
• Vaccination of health care workers in the UK reduced infection rates for their household members
• A recent study in China (with different vaccines) found vaccinated people were less likely to pass on the virus even with Delta
• An even more recent UK study looking at the Pfizer and AZ vaccines found the same

A few assorted issues:

First, the denominator questions — not the question of the actual population of NZ, which Henry Cooke is in charge of, but the eligibility and ethnicity questions.

Should we be quoting vaccination as a percentage of those eligible or as a percentage of the population? Yes, both. They answer two different important questions.  There’s a question of epidemic dynamics: are we getting to a point where enough people are vaccinated for Delta to be controllable more easily? What’s relevant there is vaccination as a percentage of the population; kids still count as unvaccinated, even though they aren’t eligible. There’s also a social question: are we providing the right access, information, and incentives to get people vaccinated? What’s relevant there is vaccination as a percentage of those eligible.

Next, ethnicity. I’ve seen people asking how ethnicity is counted in the MoH reports. Most of the NZ government tries to count people according to all the ethnicities they identify with — you can be in multiple categories. As a result, the categories add up to more than 100% of the population. The Ministry of Health does something different. If you give them multiple ethnicities, they pick one.  They prioritise: you’re Māori if that’s one of your ethnicities; you’re Pacific if that’s one of your ethnicities and Māori isn’t; you’re Asian if that’s one of your ethnicities and Māori and Pacific both aren’t, and so on. The advantage of this is that subgroups add up nicely: the number of vaccinations overall is the sum  of the numbers in each ethnic group. The disadvantage is that you may not be in the group or groups you expect.

Finally, pictures like this (this one is from @farmgeek on Twitter)

This is aiming to show the protective effect of vaccines. It’s a lot better than just reporting the % vaccinated among cases or hospitalised cases, because it shows the denominator.  The ratio of the red:green ratios in two bars is an estimate of one aspect of vaccine effectiveness; you can see it’s big.

It’s not a perfect estimate, for two reasons. The first is differences in exposure. If people who are unvaccinated are also more likely to be exposed, the vaccine will look more effective than it is; if people who are unvaccinated are less likely to be exposed the vaccine will look less effective than it is.  Both of these are likely: vaccination and exposure is broadly higher in Auckland than in the rest of the country, but within Auckland vaccination is  higher in areas where exposure is probably lower.

On top of any differences in exposure, a graph like this underestimates the impact of the vaccine because it misses out the reduction in unvaccinated cases due to the vaccine. Getting vaccinated protects you, but as the vaccination rates slowly rise, getting vaccinated also increasingly protects other people, regardless of their vaccination status.  Measles is a good example here: vaccinated people are almost never hospitalised for measles, because the vaccine protects us, but very few unvaccinated people are hospitalised for measles because community vaccination levels slow the outbreaks down enough for testing and tracing to control them.

Various people have created graphics showing the breakdown of vaccination rates across subpopulations of New Zealand.  They aren’t great (the vaccination rates, not the graphics), but they are improving.  As the graphics show, vaccination rates are lower in some subgroups than others.  Even when we get to 90% coverage on average, we could be well below 90% for some groups of people. This is a problem for two reasons.

The first reason is obvious: equity. People who haven’t been vaccinated yet aren’t just freeloading, they have reasons. For some people it’s harder to get to a vaccination (because of work hours or because they live somewhere remote). Others don’t trust the medical system — often for reasons that were well founded historically. It’s important to make sure everyone has a real opportunity to get vaccinated.

The second reason is less obvious and more statistical: we need a higher vaccination rate if the unvaccinated are not evenly distributed through society.  A cluster of people with lower vaccination rate will not only be at risk of Covid themselves, they will be an opportunity for Covid to spread. This is true of ethnic groups, but also of churches, dog-walkers, soccer moms, fans of provincial rugby, or nerds at statistics conferences.

Modelling the full complexity of NZ society and Covid dynamics is beyond what I have the data and computation resources to do, so I coded up a very simplified model to show, qualitatively, the sorts of things that can happen.  This is a fairly common use of mathematical models: not just to predict what will happen, but to show the range of behaviours that are possible.

The model is a 100×100 grid, where people can only infect their neighbours (no-one accidentally flies to Wānaka or has a job as a truck driver).  Vaccination reduces your risk of being infected, and also reduces your risk of passing on the infection.  With a random 83% of the population vaccinated, the outbreaks can’t spread far (83% of the NZ population is about 95% of the 12+ population). Here are two random outbreaks. Blue is vaccinated, grey is unvaccinated; purple is vaccinated and infected, red is unvaccinated and infected

Now, suppose we have the same 83% vaccination on average, but there’s a high-risk group (lower left) who are less vaccinated and who cluster together. If we’re lucky, a random outbreak misses them; if we’re not, it hits them

Having a non-uniform spread of unvaccinated people increases the number of cases for them, and also for vaccinated people.

We can get more dramatic sorts of clustering, where a group of unvaccinated people are connected to each other and also across society. Again, if we’re lucky, the outbreak hits only vaccinated people; if we’re not lucky, it spread very widely and more vaccinated people are infected than with a uniform spread. Do you feel lucky?

And a more dramatic example, with criss-cross connections of unvaccinated people

These obviously aren’t realistic depictions of New Zealand society, which isn’t square or blue and has lots of long-distance connections. They are, though, depictions of the sort of impact that population structure is able to have on disease spread. These example all have the same overall, high, vaccination rate, but they have very different outbreaks.

It’s not enough to get good vaccine coverage on average. Every subgroup matters.

Academic journals keep expanding, especially with the growth of open-access journals. Some publishers have reacted to this by creating a bunch of new journals. A few of these publishers give all the journals related names. Nature has Nature Genetics and Nature Communications and Nature Scientific Reports. The BMJ has BMJ Open and BMJ Nutrition, Prevention and Healthcare. The Lancet has The Lancet Public Health and The Lancet Regional Health — Western Pacific and others.

These journals are not the same as the parent journal.  You might or might not think a paper published in Nature was especially reliable because it’s hard to publish in Nature; that’s much less true for Nature Scientific Reports. You might comment that research has been published “in prestigious medical journal The Lancet“, but that’s misleading if it was actually published in The Lancet Regional Health — Western Pacific. I think the importance of journal rankings is vastly overrated, but if you’re going to rely on it you need to get it right.

Back in November, the Trump Campaign held a famous press conference that was not exactly at the Four Seasons hotel.  The label Nature Total Landscaping for these additional journals is a bit unfair — Four Seasons Total Landscaping isn’t even trying to be in the posh hotel business — but it was irresistible to science social media.