Posts written by Thomas Lumley (2534)

The idea of cutting down on lattes to be able to afford a house has cropped up again. The proximate cause is a Newshub story that doesn’t quite go there — but it does talk about rent costs and mortgage rates and about satisfying a home lender under the new CCFA credit provisions, so it’s pretty close.

Now, first, I will agree that there are almost certainly people out there who haven’t emotionally grasped that buying 200 flat whites, one per day, costs (say) $900 that you could have spent on a $900 thing instead.  I don’t know if those people are likely to be helped by the story, but maybe it’s worth a try. At the level of housing, though, $900 in a year — or even two coffees every single day, for (say) $3300 — gets you nowhere in comparison with housing price inflation.  The same is true for avocados — maybe avocado toast in a cafe costs more than a coffee, but you don’t have it every day.

You might say that coffee (or avocado) is just one example, and that the point is to pay continuous and obsessive attention to shaving the costs of everything you buy. But to keep up with the rising cost of a mortgage deposit many people would have to save more than their entire discretionary income; shaving pennies isn’t going to get you there.

Perhaps most importantly, though, these approaches can’t work for most people because the housing crisis in New Zealand isn’t due to a shortage of money to spend on housing. We’re collectively spending too much money on housing. Cutting down on coffee or avocado or any other discretionary spending, so as to put more money into the real-estate sector, isn’t going to make housing more affordable on average, even if everyone does it.

This showed up on my Twitter feed this morning, originally from here. It triggered a certain amount of wailing and gnashing of teeth from Americans.

The basic pattern looks plausible; about two-thirds of the US population vaccinated. If you look carefully at the graph, you see something else: the ‘not vaccinated’ group are broken down by attitude. This can’t be an all-ages picture: if anyone is doing large-scale surveys of attitudes to Covid vaccination among six-year-olds around the world it’s (a) a revolution in survey methods that we should know more about and (b) not all that relevant to whether the six-year-olds get vaccinated.

As the description at the link says, this was based on a survey of adults. It was supposed to be nationally representative samples of adults. It clearly wasn’t. Based on doses delivered, the USA reached 75% vaccination for adults by October; Australia is currently over 95% in adults.  The qualitative message might be true, but the numbers aren’t right.

We saw recently how two big non-random US surveys had overestimated vaccination rates, the opposite problem. Why do people do this when we already know the answer? The surveys are (potentially) useful because they ask other questions: they can break down vaccination by other attitudes and circumstance of the respondent, which the CDC data cannot do. It still matters if the answers are right, though.

• Covid was the leading cause of death “in the line of duty” for US law enforcement officers. (Also, the report giving the numbers had this graph
  
  The ‘Covid-19’ section of the bar, which should be nearly two-thirds of the total, is less than half)
• An attempt to check up on a sample of India’s forest-planting program
• Story on vaping from the Herald: it appears an Asthma and Respiratory Foundation survey gets very different results from the NZ Health Survey on the proportion of teens who vape daily.  That’s pretty concerning: 5.8% vape daily vs 20% ‘addicted’ means at least one of the surveys is badly wrong.
• People are more likely to trust pretty graphics (separately from other reasons to trust the information)
• Overestimating ‘incidental’ Covid diagnoses in London hospitals

The Herald has a story about ACC injury claims in Auckland during the lockdown.

Popular hobbies and chores like spring gardening, riding new e-bikes, lawn mowing, running and climbing ladders to paint resulted in 113,960 accident claims by Aucklanders during last year’s 107-day lockdown.

Reading on, you find that the e-bikes were involved in only one twentieth of one percent of that total (and not broken out  by new and old), but there’s a reasonable amount of detail given

There are usually two questions to ask about an ACC-based story like this one. First, are the injuries actually attributable to whatever it is the story is about? Second, is there an actual increase in injuries.  Usually the answers are “yes”, and “no, respectively: Christmas day brings injuries that are clearly Christmas-attributable, but fewer in total than a normal work day.  Usually the first question can be answered from the story, and the second takes some additional research.

This story is unusual in that we genuinely seem to have an increase in injuries, though as usual it takes some work to find this out.  Looking at historical ACC data, there were 245,149 claims for Auckland in the 2018/2019 financial year, the last year of the Before Times. Over 107 days that would scale down to about 72000, so we’re seeing nearly a 60% increase!

My guess is that the pattern would be different in other regions — it will depend on who was staying home and how risky their usual job would be.  Compared to the times before Covid, the lockdown would have increased some risks and decreased others; in Auckland the increase seems to have been larger than the decrease.

In the Herald today, Covid 19: Officials predict when Auckland Omicron outbreak will happen

They say 2-4 weeks. That looks to me like an estimate of the midpoint of the uncertainty: we’re thinking of a uncertainty distribution, and the middle is at about 3 weeks, say 2-4 weeks.  It doesn’t look like a prediction range.

The reason it doesn’t look like a prediction range is that the time until an Omicron outbreak isn’t like boiling an egg, it’s like getting in a minor car crash. Every day, there’s a small risk that Omicron will escape the MIQ controls and get out into the community.  The risk is higher (per day) than it was for Delta, because Omicron is more transmissible and because we have a lot more cases in MIQ than we ever had before. It may be lower because we may have fixed up gaps in MIQ that let cases in before (though we don’t actually know how the August outbreak started).

Lots of processes in the world work roughly like this: there’s a small chance at each point in time, but there isn’t any memory — getting through today doesn’t make tomorrow safer or less safe.  The distribution of times to the first event in a memoryless process with a median of three weeks looks like this:

If the experts are right that 3 weeks is the 50/50 bet for how long we have, the range of possibilities will be much broader than 2-4 weeks. If we’re very lucky, the outbreak could be 2-3 months away. If we’re very unlucky, it could start today. Or yesterday.

• Last June, the FDA approved a medication for Alzheimer’s Disease, although a lot of people, including their external advisory committee, thought there wasn’t convincing evidence that it works.  The Centers for Medicare & Medicaid Services, the organisation that would be paying for a lot of this treatment, have decided they need some randomised trials showing that the treatment is safe and effective, so they have said they will only fund the treatment in trials.
• According to Pew Research, “Overall, about half of U.S. adults (48%) say that most things in society can be clearly divided into good and evil, while the other half (50%) say that most things in society are too complicated to be categorized this way.” This is another example of a survey question where only differences or changes are really meaningful and we can’t straightforwardly interpret the absolute value. I mean, how about “eggplant” or (for something more obviously socially constructed) “driving on the left-hand side of the road”.  I would have said a lot of things in society have no particular moral valence, but also it’s not clear how you’d categorise “most” in cases where “good or evil or complicated” is a genuine question. The differences between groups are still potentially interesting, as might be changes over time.
• Not precisely ‘statistics’ or ‘in the media’, but a YouTube video of Adam Savage (Mythbusters) talking about measurement via Graeme Edgeler
• I gave a talk about StatsChat at a conference on undergraduate statistics and maths teaching

The governor of Florida, Ron DeSantis was widely quoted last week as saying

 “Before COVID did anyone go out and seek testing to determine if they were sick? It’s usually you feel like you’re sick and you get tested to determine what you maybe have come down with.” 

As most people will know, the answer is “yes, they did”.  You can’t get near a doctor without having your blood pressure measured, because high blood pressure is common, not obvious without testing, and treatable.  There are tests performed on infants (for genetic disorders such as PKU and cystic fibrosis), on children (vision and hearing screening), and on the middle-aged and old (cholesterol, glycated hemoglobin, cancer screening). There are tests mainly done for high-risk groups (TB, HIV); there are tests done before you get certain medications (liver or kidney function).   Until 1986, Florida required a test to see if you had syphilis in order to get a marriage license.  Governor DeSantis has since said that of course he knew all that, and it was obvious he had specifically meant daily or weekly testing for a viral respiratory infection was unprecedented, not all the normal and routine ways people seek out testing to determine if they are sick. And who knows? Maybe he did mean that.

In some ways there’s surprisingly little population screening.  Screening has traditionally been popular, because people like knowing things and having explanations.  People often argue for more or say they have discovered a way to do more. There are three big barriers in going from a test you take when you feel like you’re sick and one you give to people in advance.

The first barrier is that you have to be able to do something with the result. There’s no point in taking healthy people and finding some illness unless you can do something about it.  In the endemic-Covid testing setting,  you can isolate and not go infect your workmates or your grandparents or the blokes at the pub or whoever.   In the 2004 SARS outbreak we didn’t have rapid antigen tests, but temperature screening was used to find people who might not realise they had SARS or weren’t looking to find out for some other reason.

The second barrier is the base rate problem. The New York Times had a very good story recently about prenatal genetic testing. This looks for very rare genetic disorders, usually as add-ons when testing for Down’s Syndrome. Because those disorders are very rare, most fetuses don’t have them — even most of those who test positive.  The Times reported on a set of tests where a positive test had an 80% or higher chance of being wrong.  In one sense these tests are very accurate — a negative result is very likely to be correct– but just assuming no-one has these conditions is almost as likely to be correct and has no false positives.  The base-rate problem goes together with the problem of what to do; follow-up tests are expensive.

The base rate problem is also an issue for Covid testing here in New Zealand; in contrast to many parts of the world, we currently have very low community prevalence, so a rapid antigen test positive in someone without symptoms or known exposure would very likely be a false positive.  In Victoria, the opposite is true: the base rate is high enough that the only rapid antigen tests where PCR follow-up is recommended are those in asymptomatic people with no known exposure.

The third barrier is Rose’s Prevention Paradox: for a lot of diseases, most of the cases don’t happen in high-risk people. There are people at very high risk of heart attack (what Rose was interested in), but they account for a relatively small fraction of all heart attacks.  There are people at very high risk of premature birth, but they account for a relatively small fraction of all premature births.  Someone with a blood alcohol of 0.2 has a very high risk of getting in a crash, but most car crashes aren’t like that.

Unlike the base-rate problem, Rose’s Prevention Paradox isn’t universal.  Testing for the common CFTR mutations will pick up most cases of cystic fibrosis; the Ishihara plates pick up most cases of deficient colour vision; most lung cancer is in smokers; most liver cancer in Western countries is in heavy drinkers or people with Hepatitis B or C.

The Covid example of the prevention paradox is the recent and controversial CDC announcement that most (vaccinated) people who get seriously ill from Covid have co-morbidities.  On top of the issue of whether that’s actually a cause for rejoicing, there’s the problem that the majority people who don’t get seriously ill from Covid also have co-morbidities. That is, it’s hard to pick out high risk people.  The CDC defined ‘high risk’ very broadly, so that too  many people are ‘high risk’; if they define it too narrowly, they would miss a lot of the serious Covid cases.

Very early in the pandemic, the estimate was that a Covid case gave you roughly a year’s worth of mortality risk — if everyone got Covid over a period of one year, death rates for that year would be double what they normally are, across a wide range of subgroups.  Omicron is worse than the original strain, but the vaccine helps a lot: vaccinated people are much less likely to die or become hospitalised.  Not getting Covid helps even more; masks, ventilation, distancing, testing, etc.

According to a poll published in the Washington Post, a substantial majority of Americans think it is never justified for citizens to take violent action against the government. Given the US reverence for George Washington and others who fought in the Revolutionary War, this seems a bit strange. It’s hard to interpret what any particular percentage would mean.

What’s important about the poll is that the percentage is down compared to previous polls with the same question. It’s hard to interpret the level of agreement, but it seems pretty reasonable that a decrease indicates more willingness to consider violence against the government as an option that might come into play in the foreseeable future.  Or, at least, it would be if the polling approach hadn’t also changed, from phone to online, complicating any interpretation of changes.

Similarly, when the poll finds 30% of people claim to think there is solid evidence of widespread electoral fraud in 2020, it’s a bit hard to tell what that really corresponds to — how much is actual belief and how much is going along with a party line.  The fact that it’s about the same percentage as a year ago is more informative, as is the fact that it’s higher than similar questions about past elections.

Asking the same questions over time is a much better way to pick up changes than asking people if their opinion has changed. As a strategy, it can conflict with asking the best question, and that’s an ongoing tension in public opinion research and official statistics.

(The rugby prediction posts, while popular, are most interesting before the games actually happen: predicting the past is relatively easy)

First, the posts, regardless of year of writing, with most 2021 hits

1.  What’s a Group 1 Carcinogen? (2013) Points out that the IARC classification is not about severity or danger but about the types and amounts of evidence. Sunlight is a Group 1 Carcinogen, so are alcohol and plutonium.
2. A post about a Lotto strategy that doesn’t work(2012), as an argument about the usefulness of abstract theory. See also, the martingale optional stopping theorem
3. A climate change post about graphs that shouldn’t have a zero on the y-axis(2015)
4. From October 2020, but relevant to the news again in March this year, on crime rates in the Cuba/Courtenay area of Wellington and denominators
5. Actually from July this year, one of the StatsChat Dialogues: Q: Did you see that learning maths can affect your brain? A: Well, yes. There wouldn’t be much point otherwise

And the top 2021-vintage posts

1. Number 5 from the previous list
2. From October, on interpreting vaccination percentages
3. From April, why there’s so much fuss about very rare adverse reactions to vaccines (the AZ blood clots)
4. From October, why population structure matters to epidemic control, aka, why we need to vaccinate every subgroup. Has pictures!
5. From June, how a cap-and-trade system for (a subset of) emissions messes up our intuition about other climate interventions.

These are WordPress page views: their relationship to actual readership is complicated; keep in a cool, dry place away from children; may contain nuts.

A gentleman who is Not New Zealand’s Favourite DJ has tested positive for Omicron on his 9th day in NZ, after prior negative tests.  It seems surprising that a positive test could take so long — one theory is that it’s the sort of sporadic positive you can get for a while after recovery.  On the other hand, it’s worth thinking about why it’s surprising.  New Zealand keeps seeing strange Covid occurrences: long incubation period, transmission from very brief contact, and so on. Why us? It’s us because no-one else would be able to tell.

The normal assumption if a London DJ tests positive for Omicron on December 25 is that they got it in London a few days earlier. In this example, ‘a few days earlier’ means Waiheke Island and however he got there from MIQ;  in contrast to almost everywhere else in the world, Omicron isn’t circulating in Auckland.  The week before that he was in MIQ, so the next conclusion is that he got it there; unfortunate but not unprecedented. Without genome sequencing we’d stop there, but his viral genome doesn’t match any of the three he could potentially have picked up in MIQ.  We’re now down to weird possibilities, and the least weird is that he was carrying it all the time. But if he was almost anywhere else in the world, we wouldn’t even be starting to think about the weird possibilities.

It’s fairly easy to estimate the low end of the time-to-positive-test distribution: someone goes to a party; a few days later there are twenty cases; you do the maths. To get the low end you need some cases where you’re sure they couldn’t have been infected before a specific exposure, because they weren’t exposed before then. At the start of an outbreak that’s fairly easy. To get the high end of the distribution you need some cases where you’re sure they couldn’t have been infected after a specific exposure. That’s a much less common scenario, so the data aren’t as good.