Posts written by Thomas Lumley (2646)

From USA Today, via Twitter

These figures don’t make a lot of sense.

Firstly, that’s a really high proportion of median income.  You can easily find median household income in the US as just over $60,000.  If a typical household had two members who each spent $18,000 on “nonessentials”, that’s more than half their total income, which is hard to believe.  [We’re just doing obvious credibility checks here, but from more detailed data (table H-15)  the median per-person household income is $34,000]

It’s still possible that the figures are true, but really are means and have nothing to do with the “typical American”.  But consider the numbers for subscription boxes and rideshares. If these are means for adults in the US, we can multiply by 250 million to get the corresponding totals of just under and just over $290 billion per year. A quick Google search finds that Uber’s gross income from bookings was about $14 billion last year (which I think includes Uber Eats as well as rideshare). Uber is the largest rideshare company, and there’s no way it’s only 5% of the total market.

It’s harder to find data for e-commerce subscription boxes, but McKinsey says total revenue for the companies in that category that are on a list of the top 500 internet retailers was $2.6 billion in 2016, and whatever people say about long-tail e-commerce, I don’t believe those top companies have less than 1% of the market.

If you click through to look for the data source, there’s nothing there that gives any information. Nothing about the questions asked; nothing about how the people were selected; nothing about how representative the sample ended up being.

Chris Knox (of the NZ Herald) pointed me to a twitter thread where Kate Rabinowitz (of the Washington Post) compared the numbers to the Bureau of Labor Statistics’ Consumer Expenditure (CE) Survey.  Clearly some reporters understand the issue.  But it shouldn’t be possible for a life insurance company to report a dubious survey as a way to get advertising in a major newspaper. Make them buy an ad instead.

From the Telegraph

Sportswomen should go on the pill to avoid career-ending injuries, the authors of a groundbreaking study have suggested.

Analysis of more than 165,000 female athletes in the US found those taking the oral contraceptive were significantly less likely to suffer knee injuries.

There’s a few problems with this as a summary of the research(press release)  The study didn’t measure knee injuries — it measured ACL surgeries. More importantly, it wasn’t of female athletes, but of any women aged 15 to 49 in a big medical-records database who weren’t pregnant and didn’t have a known connective tissue disorder — the study didn’t have any data at all on physical activity. Also, ‘groundbreaking’ seems a bit much given that the research paper talks about previous studies with similar results, as well as biochemical reasons that there might be a relationship.

Perhaps the most problematic aspect of the Telegraph story, though is how it relates to this graph from the research paper. Basically the whole association is driven by the 15-19 age group.

This is obviously an important research question in sports medicine, and of public interest, but that makes it more important to get things right. The Guardian and the Miami Herald did a lot better at describing who was studied and the limitations of the results — such as not assessing physical activity at all.

Stats New Zealand have released the first full quarter of real-estate transfer data since the introduction of restrictions on purchases by non-residents (except for Australians and Singaporeans).

There are few stories talking about the fall in sales.  The data include residence information on the seller as well as the buyer, so I made some graphs showing how the net transfers have changed over time.  In all these graphs, the orange line is the number of buyers in the group, the blue line is the number of sellers, and the black line is the difference: the flow of real estate towards the group

First, citizens.

Now citizens or people with residence visas. There’s almost perfect cancellation, but a slight flow of properties towards people identified as citizens or residents — not only this quarter, but for the whole time range.

 On this scale, non-residents are basically invisible  Zooming in by a factor of ten, the net acquisition by non-residents was positive until this quarter, but roughly half the gross numbers that were usually quoted. This quarter it was negative.

 Some buyers or sellers had unknown residence status, though the numbers fell markedly after the start of 2018. Adding these to the non-residents doesn’t make a big difference.

Residence isn’t well-defined for corporate buyers — you could be overseas and own an NZ trust.  Some people add all this group to the ‘non-residents’.  We now have to zoom back out to the original scale. Interestingly, the net transfer here is negative — if you’re the sort of person who thinks overseas ownership is bad, treating all corporate ownership as foreign isn’t any sort of conservative or ‘worst-case’ analysis. 

Yesterday, the Police released the first data from the expanded program of wastewater drug testing they have been doing with ESR (more detail).  There’s a report at Stuff (from the Dominion Post).

The data are estimated total consumption of various drugs by the people using various parts of the sewer system in New Zealand.  There’s obviously quite a bit of uncertainty, but it’s probably a useful guide.  Because they’re measuring metabolites, not the original drugs, they are less likely to be misled by drugs discarded without being consumed — but they are probably more sensitive to differences in metabolism between people.  The analogue for tobacco and alcohol is the data on total amount sold, which is a widely used statistic because it’s thought to be more accurate than self-report.

The headline figure is an estimated 16kg of methamphetamine used per week. So, how does this compare to other estimates?  The story at Stuff has part of the calculation, showing about 1000 mg per day per 1000 people in the highest-testing areas, and a national average of about 600 mg/day/1000 people.  If NZ survey-based statistics are accurate, perhaps 2 people per 1000 will be regular users, so we’d be looking at 300mg/user/day. That’s a lot, but it’s within the range that Erowid lists for chronic users.  So, probably the surveys are underestimating the number of users, but not by a factor of ten or anything like that.

As Russell Brown says in his analysis of the report (which you should definitely read), one of the important findings is that total meth use seems to be pretty constant day to day, unlike MDMA use, which peaks at the weekend.  That’s consistent with a subset of regular users taking it daily or most days.  One of the limitations of wastewater testing is that the social impact of a drug doesn’t just depend on the total community consumption, but combining the wastewater data with other data is helpful.

The comparison of fentanyl to the other drugs is confounded with the higher potency of fentanyl — something it’s unusual for the media to gloss over.  If we go by the therapeutic dose of fentanyl for post-operative pain, 50-100 micrograms, there are about as many doses in the 3g of fentanyl reported per week as in the 4kg of MDMA.  On the other hand, the 3g does include prescription use, and people who were dependent on the drug would likely take higher doses and multiple doses per day.

Yesterday, the government released results from the 2018 PBRF round.  PBRF is designed to allocate research funding to tertiary institutions.  The idea is to do this in a way that more money goes to institutions that do more research, both so that the status quo is upheld in the short term, and so that institutions such as AUT can make progress and get a larger share in the medium term. There’s also a goal of providing coarsely-grained feedback to institutions on where their research has been most successful. You could disagree with the goals and argue that all funding should be based on competitive grant proposals, or that there should be more research funding given to institutions that involve under-represented groups in research, or that people who don’t get good research scores should be fired, or whatever, but I think PBRF does reasonably well at achieving the goals it has.

An unfortunate side-effect of PBRF is that the government produces a fairly large set of numbers about the tertiary institutions.  Given a large set of numbers about a small number of institutions, there’s a very high chance that you can find a somewhat sensible summary of those numbers that puts you on top.  This is a game everyone can play. And does.

You don’t even need to go past the TEC information release. As Radio NZ says

The report said Victoria University had the highest score (29.19) for the number and quality of active researchers on its teaching and research staff, followed by Otago (26.09) and Canterbury (25.92) universities.

However, Lincoln University had the highest score for active researchers compared to the number of students enrolled in courses at degree level and higher (2.14), followed by Otago (2.06), and Catholic theological institute the Good Shepherd College (2.00).

The results showed the largest university, Auckland, had 1744 FTE staff funded by the PBRF and 391 or 22 percent were classified as A grade or world-class, the highest proportion of any university.

At Victoria University 20 percent or 173 of 865 funded-researchers received As, while at the University of Otago the figure was 17 percent or 229 out of 1358 researchers.

Moving further out, AUT has the fastest relative increase in funded researchers; Massey has a 40% increase in those at the top level; and Universities New Zealand astonishes everyone by pointing out universities did a lot more research than polytechs and wānanga.

The problem with these rankings …. the two problems with these rankings…. among the problems with these rankings are the fineness of the distinctions needed (well below the precision of the measurement process), the way everyone suggests the ranking they do best on is most important, and the effort wasted on tuning the system to target particular rankings.  Perhaps most importantly, though, there’s a corrosive effect of university leaders pretending that something is important and quantitatively well-founded when they know it isn’t, or at best, know they don’t have good evidence that it is.

Fred Vultee, an academic and former editor has written about an analogous problem for newspapers

It’s a free country, and Congress shall make no law abridging, &c &c &c, so you’re perfectly entitled to do whatever you want with your “analysis” time and your frontpage space. In some cosmic sense, sharing popular delusions about the lottery isn’t any worse than publishing the horoscopes. But if you’d like to be taken seriously when you proclaim “study says” or “poll reveals,” you need to run a disclaimer with your “want to hit the lottery” tales

• A good use of computerisation in criminal justice: efficiently expunging non-violent cannabis-related convictions in California
• Missing Numbers is a blog about the data that the government should collect and measure in the UK, but doesn’t.
• 3-D visualisation of flight patterns around major airports
• “While AI can be an incredibly useful tool, we should proceed with caution until machines achieve a sufficiently nuanced understanding of human values that they do not become obsessed with constructing an armed compound on their own private island.” From America’s Finest News Source
• Beautiful visualisation about migrants in Brussels (one of the few places with a greater proportion of people born outside the country than Auckland)
• The power of effective (and, arguably, misleading) maps

Imagine if the U.S. prison population had six senators. https://t.co/cIseBxRO0A

— David Slavick (@davidslavick) April 24, 2019

A few years ago I briefly ran a Twitter account that retweeted science news stories with, as it might be, “In mice”, “In rats”, “In cell cultures”.

Now someone else has a new (and hopefully longer-lasting) effort: @justsaysinmice. For example

and

The point isn’t that there’s something wrong with mouse research. These types of basic science can be enormously valuable. The problem is the headlines (and ledes and first few paragraphs and…) which make much stronger claims about health than the research supports.

Someone seeing “Reused cooking oil ups risk of metastases in breast cancer patients” should be able to assume that that’s what the study found — or even without the causal language, that the study involved metastases in breast cancer patients, rather than  tumours injected into mice.

Scientists should be objecting when their work is mischaracterised that way, but they mostly don’t. Some of them even encourage it.  Readers should also object, but again, they don’t and many of them encourage it.

• How IBM Watson Overpromised and Underdelivered on AI Health Care. IEEE Spectrum
• “In a major ethical leap for the tech world, Chinese start-ups have built algorithms that the government uses to track members of a largely Muslim minority group”. New York Times
• A description of the process behind a major data visualisation “The Design Process of “Why Do Cats & Dogs …?”
• I’ve written about the University of Washington course “Calling Bullshit” before, but it’s worth reading the writeups in Pacific Standard and  the Guardian
• Following the stories about identifying a serial killer using data from recreational genotyping, BuzzFeed reporter Peter Aldhous tried to track down ten of his colleagues from DNA reports. He got six of them. Of course, this is a more limited frame of candidates than in the forensic case.
• With better measurement of teens’ “screen time” you find less evidence that it’s associated with mental health.
• The Excellence in Statistical Reporting Award from the American Statistical Association goes to Hansi Lo Wang from NPR for his reporting on next year’s US Census.
• Ongoing note:  I’m on the External Data Quality Panel for the NZ 2018 census, so I won’t be commenting on the conduct or reliabilty of the census until we release our report.

From 1News, via Richard Hills on Twitter, a graphic accompanying their latest opinion poll results.

If the picture of Parliament isn’t supposed to have some quantitative relationship to the opinion figures, it shouldn’t be in party colours.

I bet you they won’t play this song on the radio
I bet you they won’t play this new (bleep) song
It’s not that it’s (buzzer) or (horn) controversial
Just that the (bell)-ing words are awfully strong

— Monty Python, I Bet You They Won’t Play This Song on the Radio

Last year, I missed mentioning one of the minor jewels of the NZ regulatory landscape: the report ‘Language That May Offend in Broadcasting’ from the Broadcasting Standards Authority.

There’s pretty wide agreement around the world that some words should be avoided in radio and TV, and that what language is appropriate would vary according to the setting.  A gritty crime drama would be allowed more license that the main news bulletin, and someone using “Jesus” to refer to the Christian messiah would be different from someone using the word just as an expletive. There’s often reference to “community standards”.  In a lot of places, though, “community standards” are just pulled out of the air or handed down in precedent.

The Broadcasting Standards Authority does a survey of New Zealand every few years, where they actually ask a representative sample of the population how they feel about various words in various contexts (and asks them to suggest other offensive words that weren’t listed).

This time, the BSA added some words or phrases from Samoan and Te Reo Māori. They also added a few English words that are pejorative references to particular groups — in contrast to religious references and content-free profanity, this sort of word has remained unacceptable or perhaps become more unacceptable.