Posts written by Thomas Lumley (2563)

Q: Did you see that ibuprofen prevents dementia?

A: No

Q: Common 3p anti-inflammatory pill ‘can help prevent dementia’ says new study.  And Prolonged Use Of Anti-Inflammatory Pills May Help Prevent Dementia: Study. And…

A: No

Q: Those quotes in the first headline imply someone really said it

A: Well, that used to be the convention, but apparently not in this case.

Q: Mice?

A: No, not mice. People

Q: Correlations, or randomised trials?

A: Correlations

Q: Is there a paper?

A:  There is a paper.

Q: Has anyone looked at this correlation before?

A: Oh my yes. So many people. There are even randomised trials. The Cochrane Review says “no evidence that it works, and it may be dangerous”

Q: What do the scientists say about that?

A: The trials were too short to show the effect, or they started too late in the disease process.

Q: Is that possible?

A: Sure, it’s possible. Lots of things are possible.

Q: So you don’t believe the paper?

A: It’s a perfectly reasonable paper for consenting scientists in private (I know some of the authors), but it doesn’t justify the recommendation of long-term treatment with these drugs for the prevention of dementia, especially given the potential adverse effects

Q: Well, that’s your opinion. What do the scientists who wrote the paper say?

A: Our results.. do not justify the recommendation of long-term treatment with NSAIDs for the prevention of dementia, given its potential adverse effects. 

“Bowel cancer does not discriminate on race. Māori and Pacific peoples have a similar risk of developing bowel cancer compared to other population groups at a given age.

“It was true that a higher proportion of bowel cancers occur in Māori and Pacific peoples at a younger age, but that is because the overall demographics of those groups are younger. It has always been age that determines bowel cancer risk, not race”

I thought that this required more explanation. Proportion at an age does not seem to me to depend on the demographics of those ages

It’s a good question.  The minister is plausibly correct that the younger age at diagnosis for Māori and Pacific Peoples is driven substantially by demographics.   That is, one reason that few very elderly Māori get diagnosed with bowel cancer is that there aren’t many very elderly Māori.  Fewer Māori make it to be very elderly, and also the Māori population is increasing both because of higher birth rate and because more people are now identifying as Māori to the healthcare system.   If you have the same age-specific incidence rates in two groups, but one group averages younger, the incidence in the younger group will tend to be at younger ages.

That’s not the whole story, or even most of it, for this decision.  The age distribution being quoted is age at diagnosis (because that’s what you have for cancer), but diagnosis isn’t equally distributed. According to the HQSC atlas of health care variation, Māori are more likely to be diagnosed with bowel cancer following emergency presentation, and more likely to die within two years of diagnosis. This means there’s a good case that Māori are more underdiagnosed than New Zealand Europeans.   On the other hand, a lot of the problem with underdiagnosis is not being tested. If people end up just doing one test when they become eligible, and then not bothering later, the optimal age for a single test is more like 60 than 50.   And, as a general principle,  if you want to spend public money to improve health, you typically will do better by spending it on people who are less healthy and have worse access to health care.

In any case, while the headline Bowel cancer screening eligibility age to be lowered for everyone is technically true (since the lower-age screening pilot has already been stopped), it seems to miss the point a little.

NZ$ Herald has a piece today about increases in methamphetamine consumption (as measured by wastewater)

The story is about cheap meth imports increasing the supply, with a speculative aside into a decrease in consumption in Ōpōtiki and whether it might be attributable to major gang arrests.

The sudden spike in 2024 is surprising. Normally I would guess that sudden level changes in a time series were due to different definitions of the variable, but it appears to be defined the same way.

It’s obviously tempting to attribute the sudden change to the relaxation of rules on pseudoephedrine sale in 2024, and some people on social media seemed happy to do so.  There are problems, though.

First, the suddenness is still strange. You’d expect the gangs to take a while to ramp up production and sale after years out of the business — and if they were ready on day one, the spike would have been in June rather than July.

Second, that’s a big spike.  There’s about gram and a half of pseudoephedrine in a pack of decongestant tablets, which will make under a gram of meth, so a 20kg jump in meth production is a lot of Sudafed: tens of  thousands of packets per week of a drug sold only by pharmacist prescription. Someone should have noticed.

The Drug Foundation says it’s due to international supply, which seems plausible — we’re seeing  300-500kg seized most quarters recently, which is 15-40kg/week, so unless we’re catching the majority there will be enough getting through to explain wastewater levels.

The Herald also quotes price data from the Drug Trends Survey, which says prices are down. That’s a bit less supportive than it looks. First, prices were only down in roughly a straight line over several years. Second, the say they think demand has saturated, which doesn’t fit the spike. And finally, the last edition of the survey sampled from February to July 2024, so mostly missed the spike

• Petrol price margins are up: RNZ.  These, as we’ve discussed, are based on the difference between the import price and retail price, and include transport and vendor costs (like running petrol stations) as well as profits.  The change from week to week in the margin is still a good indicator that either the efficiency of the system has gone down or the profits have gone up.
• The Trump administration has cited a statistic that only 6% of federal government workers are full-time in their offices. This is facially implausible when you consider how many federal jobs can’t possibly be done remotely. Pro Publica looked into the number, which came from a self-selected online survey of 6338 people claiming to be federal employees, run by a a small news organisation called Federal News Network in the suburbs of Washington.  Say no to statistics from bogus polls!
• The UK is running a big (700,000 women) trial of “AI” to assist in reading mammograms. This isn’t ChatGPT, it’s from earlier generations of image-analysis software that actually works.  The hope is that using software plus one radiologist is at least as good as the current standard of two radiologists, allowing more screening to be done without needing more scarce radiologists. The media coverage is nearly all positive and based on the government’s publicity release, eg, this from the BBC. There’s not a lot of information available about how they will measure success or exactly which “AI” platforms they will use. I couldn’t find a trial registration, though I did find one for a much smaller trial 15-20 years ago on the same question.  One negative piece of media coverage from the New Statesman, for balance.
• A US trial of an oral Covid vaccine, hoping to provide both better protection and easier delivery, has been stopped by the US government just before it was due to start. At the moment this is being described as a ‘pause’ rather than permanent stoppage.

• USA: Unfortunately, the Economy Runs On the Data Trump Is Trying to Delete (article, podcast discussion)
• New Zealand: Len Cook: Trust must be recovered after a second Census failure (The Post)

As you have probably heard, a modest-sized asteroid will pass really quite close to the earth in seven years. Perhaps very close. Maybe even closer than that. If this happens, it won’t be civilisation-destroying but could be locally devastating. The main StatsChat relevant issue is that the probability of the asteroid hitting the earth keeps going up: 1.4%, 2.2%, 3.1%.  Why is it getting consistently higher? Does that mean we can expect it will keep going up?

Most likely, the probability will keep going up for a while, then suddenly plunge to zero — though it might keep going up and up to 100%.  The sudden plunge to zero and the increase before that point happen for the same reason.    Because of our limited measurements so far, we know only approximately where the asteroid will be in seven years’ time.  You can think of a little moving cloud of possible asteroid positions in space. At the moment, the cloud of possible positions intersects where the Earth will be, but the Earth is a pretty small target (Space is big. Really big.)

As we get more precise information (for example, from the Webb Space Telescope) the cloud of possible positions gets smaller.  If the cloud still intersects the Earth’s position, the smaller cloud of possibilities means the probability of impact goes up.  If it doesn’t still intersect the Earth’s position, the probability of impact drops to zero.

If the randomness in predictions is approximately correct in some consistent way, the predicted average value of the estimate should stay the same over time (the technical term is ‘martingale’). The chance that the estimate has gone to zero will go up over time, so the estimate in scenarios where it doesn’t go to zero must also go up over time.

Update: it’s down to 0.36%
Update: and now down to 0.0022%. Asteroid goneburger.

When you get access to some data, a first step is to see if you understand it: do the variables measure what you expect, are there surprising values, and so on.  Often, you will be surprised by some of the results. Almost always this is because the data mean something a bit different from what you expected. Sometimes there are errors. Occasionally there is outright fraud.

Elon Musk and DOG-E have been looking at US Social Security data. They created a table of Social Security-eligible people not recorded as dead and noticed that (a) some of them were surprisingly old, and (b) the total added up to more than the US population.

That’s a good first step, as I said. The next step is to think about possible explanations (as Dan Davies says: “if you don’t make predictions, you won’t know when to be surprised”). The first two I thought of were people leaving the US after working long enough to be eligible for Social Security (like, for example, me) and missing death records for old people (the vital statistics records weren’t as good in the 19th century as they are now).

After that, the proper procedure is to ask someone or look for some documentation, rather than just to go with your first guess.  It’s quite likely that someone else has already observed the existence of records with unreasonable ages and looked for an explanation.

In this case, one would find (eg, by following economist Justin Wolfers) a 2023 report “Numberholders Age 100 or Older Who Did Not Have Death Information on the Numident” (PDF), a report by the Office of the Inspector General, which said that the very elderly ‘vampires collecting Social Security’ were neither vampires nor collecting Social Security, but were real people whose deaths hadn’t been recorded.   This was presumably a follow-up to a 2015 story where identity fraud was involved — but again, the government wasn’t losing money, because it wasn’t paying money out to dead people.

The excess population at younger years isn’t explained by this report, but again, the next step is to see what is already known by the people who spend their whole careers working with the data, rather than to decide  the explanation is the first thing that comes to mind.

• CDC’s “disease detectives” halved as part of DOGE cuts at health agencies
• Bird flu seen in cow veterinarians
• Louisiana will no longer promote mass vaccination (paywalled original)
• Measles outbreak continues to expand in Texas

The Herald (from the Telegraph) says People with divorced parents are at greater risk of strokes, study finds

The study is here and the press release is here. It uses 2022 data from a massive annual telephone survey of health in the US, the Behavioral Risk Factor Surveillance System, “BRFSS” to its friends.

Using BRFSS means that the data are representative of the US population, which is useful.  On the other hand, you’re limited to variables that can be assessed over the phone.  That’s fine for age, and probably fine for parental divorce.  It’s known to be a bit biased for BMI and weight. The telephone survey doesn’t even try to collect blood pressure, cholesterol, or oral contraceptive use, all known to be risk factors for stroke.  And if you call people up on the phone and ask if they’ve ever had a stroke, you tend to miss the people whose strokes were fatal or incapacitating (about a quarter of people die immediately or within a year if they have a stroke).

Still, the researchers collected some useful variables to try to adjust away the differences between people with and without divorced parents.  As usual, we have to worry about whether they went too far — for example, if the mechanism was via diabetes or depression, then adjusting for diabetes or depression would induce bias in the results.

This sort of research can be useful as a first step, to see if it’s worth doing an analysis using more helpful data from a study that followed people up over time — either a birth cohort study or a heart-disease cohort study.  It’s interesting as initial news that there’s a relationship — though you also might think adverse effects of divorce would get smaller in recent decades as divorce became less noteworthy.

All this is background for my main point.  While looking for links to published papers, I found that one of the same researchers had done the same sort of analysis with the BRFSS data from 2010 and published it in 2012. They found a stronger association twelve years ago than now.  I don’t know about you, but I would have appreciated this fact being in the press release and in the news story.

Listening to the Slate Money podcast, I heard about an interesting survey result

Elizabeth Spiers:  My number is 73 and that’s percent and that’s the number of men in a recent YouGov survey who say they do most of the chores in their household.

I found a Washington Post story

60 percent of women who live with a partner say they do all or most of the chores. But 73 percent of similarly situated men say that they do the most — or that they share chores equally.

Here’s the top few lines of the YouGov table

Breaking out advanced statistical software, 23%+19% is 42%, and 33%+30% is 63%. The figure for women matches the story, allowing for reasonable rounding. The figure for men doesn’t?

If we add in “shared equally”, which is given as 33% for men and 22% for women we can get to 75% “all” or “most” or “equally” for men, but 83% “all” or “most” or “equally” for women.  And while the story is supposed to be that men say they are doing more and are delusional, the reported table has more work self-reported by women than men at all levels. It’s still possible, of course, that some the 42% of men claiming to do all or most chores are not fully aware of the situation, but the message that makes the results headline-worthy is not in the data.

The Washington Post story is still well worth reading — it uses the YouGov poll as a hook to discuss much more detailed data from the American Time Use Survey, which has the advantage that people write down contemporaneously what they are doing for an actual two weeks rather than trying to guess at an average.