Posts from January 2020 (20)

There’s a revolutionary hypothesis about the cost of US healthcare: that it’s driven by the extreme, chronic costs of a few people.  These people have problems that go far beyond medicine, but they could be helped by high-level, individualised social services and support, that would still be cheaper than their medical costs.  Atul Gawande wrote a famous article in the New Yorker on the topic.

In Camden, New Jersey, a program of this sort has been in place for several years.  A new research paper says

The program has been heralded as a promising, data-driven, relationship-based, intensive care management program for superutilizers, and federal funding has expanded versions of the model for use in cities other than Camden, New Jersey.To date, however, the only evidence of its effect is an analysis of the health care spending of 36 patients before and after the intervention and an evaluation of four expansion sites in which propensity-score matching was used to compare the outcomes for 149 program patients with outcomes for controls

What we need here is a randomised trial.  Impressively, given that the program has been famous, and has already been expanded into other cities, the Camden Coalition did a randomised trial, comparing the promising, data-driven, relationship-based, intensive care management program to treatment as usual.  As the New York Times reports

• Reuters graphics showing the size of the Australian bushfires
• From Jen Hay (linguist) on Twitter Karl du Fresne recently wrote: “The year just passed was notable for the supplanting of the letter T by D in spoken English, so that we got authoridy in place of authority, credibilidy instead of credibility, securidy for security, and so on”. Professor Hay goes on to point out that this is a long-standing trend in NZ English, and show that du Fresne himself was doing it as long ago as 2013!
• On cancer hype (via Derek Lowe) By email or through their press representatives, STAT asked 17 of the leaders who were quoted in that press release to reflect on what the moonshot has and hasn’t accomplished in the past four years. None of them agreed to comment.
• Also from Derek Lowe, about cancer trends.  Some cancers are occurring at the same rate they used to, but people aren’t dying from them (good). Some are occurring less often than they used to (good). Some are occurring much more often than they used to, but no more people are dying from them.  That sounds as though it might be good, but (at least in part) it will be overdiagnosis — the cancers aren’t really getting that much more common, we’re just diagnosing harmless cases.
• 538 are giving forecasts for the US primary elections — including uncertainty intervals, which are really wide still. In 80% of simulations, [Biden] wins between 5% and 45% of the vote. He has a 3 in 10 (30%) chance of winning the most votes, essentially tied with the second most likely winner, Sanders, who has a 3 in 10 (28%) chance.

Morning Consult and Politico run a survey where they asked registered US voters to point Iran out on a map.  About a quarter could. Here are the complete results

As you will notice, there are dots everywhere. Morning Consult didn’t go in for any grandiose interpretations; they just presented the proportion of respondents getting the answer correct. Other people were less restrained. There was much wailing and gnashing of teeth over this map, on Twitter, but also at other media outlets.  For example, Rashaan Ayesh wrote at Axios

While the Middle East saw definite clustering, some respondents believed — among dozens of wild responses — that Iran was located in:

• The U.S.
• Canada
• Spain
• Russia
• Brazil
• Australia
• The middle of the Atlantic Ocean

The claim that some respondents believed Iran to be in the US or the middle of the Atlantic ocean is worrying to me.   I can’t see how a sensible journalist could possibly state as a fact that someone had a belief like that based just on the incredibly flimsy evidence that there’s a dot there on the map.  I’d want to at least have someone explicit claim that they thought Iran was in the middle of the ocean, and ideally have follow-up questions asked about where they think Iranians keep their mosques and carpets and where they grow their rice and barberries and pomegranates and walnuts and so on.

It’s a well-known phenomenon that people don’t always give the answers you want on surveys. Scott Alexander has written about the ‘Lizard-man constant’, and that’s even before you give people a clicky interactive map to play with.   It’s barely conceivable — no, actually it isn’t, but let’s pretend it is — that some people think that large continent on the upper left of the map is the Middle East, or recognise it as North America but believe Iran is in the US.

It seems much more likely to me, though, that

• they don’t know where Iran is and would rather give an obviously wrong answer than look as if they’re trying
• they clicked wrong on the map, because interactive maps are actually kind of a pain to use, especially on a phone.
• they think having heard of Iraan, Texas (pronounced Ira-an), or Persia, Iowa, will make them look clever.
• they want to mess up the survey because they hate polling or for political reasons or because they’re having a bad day

The next question, of course, is how much it matters that the majority of US voters couldn’t find Iran on a map.  Iran is relatively easy, as countries go — I knew that it had a coastline on the Persian Gulf and that it wasn’t on the Arabian Peninsula, and Iran is big enough that this is sufficient.  But suppose I thought it was where Iraq is, or Syria, as many people did. Should have an impact on my political views (assuming I’m in the US)?  It’s not clear that it should.  The potential ability of Iran to close the Straits of Hormuz (and the Doha and Dubai airports) does depend on its location, but the question of whether the US was justified in killing Qasem Soleimani or threatening to bomb Iranian cultural sites doesn’t seem affected.

There’s a story by Brittany Keogh at Stuff on misconduct cases at NZ universities (via Nicola Gaston).  The headline example was someone bringing a gun (unloaded, as a film prop). There were 1625 cases of cheating.

As with crime data, there are data collection biases here: whether or not things are reported to the university, and whether the university takes any action, and whether that action ends up as a misconduct record, and how that record is classified.   Notably,

Victoria University of Wellington was the only university that noted disciplinary cases for sexually harmful behaviour, with three incidents reported in 2018. 

The university defined sexually harmful behaviour as “any form of unwelcome sexual advance, request for sexual favours, and any other unwanted behaviour that is sexual in nature”, including sexual harassment or assault.

There’s no way there were only three cases at universities. Or only three cases reported to universities.

The under-reporting is not quite as bad as that: the University of Canterbury reported ‘several’ harassment cases that resulted from a Law Society review into sexual misconduct, so we’ve got a classification problem as well.  Some of the harassment cases at other universities might also be included.  And there’s no data from Otago (which hasn’t responded) or Massey (which refused).

It’s definitely possible to go too far the other way — in the US, Federal law requires reporting and investigation for all incidents, even in the absence of a complaint, which means that a victim who doesn’t want to be put through an investigation is quite limited in who they can talk to.

With these numbers, though, the big story shouldn’t be that someone once brought an unloaded gun to campus with no violent intent; it should be that the universities are managing not to notice sexual harassment and assault.

There are lots of maps going around of the Australian fires. Some of them are very good, others indicate ways you can accidentally or deliberately mislead people.

This one was created as a 3-D visualisation by Anthony Hearsey, based on infra-red hotspot data accumulated over the fire period. It has been circulated as a ‘NASA photo’, which it isn’t.

If you think of it as a map or photo, the image exaggerates the scale of the fires by accumulating burning areas over time, and by the ‘glow’ effect in the 3-d rendering.  As Nick Evershed, a Guardian data journalist, points out, it is also exaggerated because it ’rounds up’ each spot of fire into a whole map grid square.  Nick produced this version with the accumulated fires over time but without the rounding and glow.

A much more effective way to mislead with maps, though, is just to make stuff up and not worry about the facts.  Yesterday, ABC News (the US one, not the Australian one) gave this map (insightful annotation from Buzzfeed)

They now have this one. It still make the fire areas look larger than they are, but in standard and hard-to-avoid ways.

A note at the end of the story now says

Editor’s note: The previous graphic in this story has been updated to reflect the hot spots around Australia.

Maybe it’s a matter of taste, but I don’t think that goes nearly far enough towards admitting what they did.

Q: Did you see that going to the opera makes you live longer?

A: No, it just makes it feel longer

Q: ◔_◔

Q: This story, from the New York Times.

Now, there is evidence that simply being exposed to the arts may help people live longer.

Researchers in London who followed thousands of people 50 and older over a 14-year period discovered that those who went to a museum or attended a concert just once or twice a year were 14 percent less likely to die during that period than those who didn’t.

A: Actually, if you look at the research paper, they were just over half as likely to die during the study period: 47.5% vs 26.6%. And those who went at least monthly were only 60% less likely to die: 18.6% died.

Q: You don’t usually see the media understating research findings, do you?

A: No. And they’re not.  The 14% lower (and 31% lower for monthly or more frequent) are after attempting to adjust away the effects of other factors related to both longevity and arts/museums/etc

Q: You mean like the opera is expensive and so rich people are more likely to go?

A: Yes, although some museums are free, and so are some Shakespeare, etc,

Q: And if you can’t see or hear very well you’re less likely to go to opera or art galleries. Or if you can’t easily walk short distances or climb stairs?

A: Yes, that sort of thing.

Q: So the researchers didn’t just ignore all of that, like some people on Twitter were saying?

A: No. The BMJ has some standards.

Q: And so the 14% reduction left over after that is probably real?

A: No, it’s still probably exaggerated.  Adjusting for this sort of thing is hard.  For example, for wealth they used which fifth of the population you were in. The top fifth had half the death rate of the bottom fifth, and were five times as likely to Art more often than monthly.  For education, the top category was “has a degree”, and they were half as likely to die in the study period and 4.5 times more likely to Art frequently than people with no qualifications.

Q: “Has a degree” is a pretty wide category if you’re lumping them all together.

A: Exactly.  If you could divide these really strong predictors up more finely (and measure them better), you’d expect to be able to remove more bias.  You’d also worry about things they didn’t measure — maybe parts of the UK with more museums also have better medical care, for example.

Q: But it could be true?

A: Sure. I don’t think 14% mortality rate reduction from going a few times a year is remotely plausible, but some benefit seems quite reasonable.

Q: It might make you less lonely or less depressed, for example

A: Those are two of the variables they tried to adjust for, so if their adjustment was successful that’s not how it works.

Q: Isn’t that unfair? I mean, if it really works by making you less lonely, that still counts!

A: Yes, that’s one of the problems with statistical adjustment — it can be hard to decide whether something’s a confounding factor or part of the effect you’re looking for.

Q: But if people wanted to take their kids to the museum over the holidays, at least the evidence is positive?

A: Well, the average age of the people in this study was 65 at the start of the study, so perhaps grandkids.  Anyway, I think the StatsChat chocolate rule applies: if you’re going to a concert or visiting a museum primarily for the health effects, you’re doing it wrong.

Team Ratings for Round 9

The basic method is described on my Department home page.
Here are the team ratings prior to this week’s games, along with the ratings at the start of the season.

Performance So Far

So far there have been 48 matches played, 36 of which were correctly predicted, a success rate of 75%.
Here are the predictions for last week’s games.

Predictions for Round 9

Here are the predictions for Round 9. The prediction is my estimated expected points difference with a positive margin being a win to the home team, and a negative margin a win to the away team.

Team Ratings for Round 8 Delayed Match

The basic method is described on my Department home page.
Here are the team ratings prior to this week’s games, along with the ratings at the start of the season.

Performance So Far

So far there have been 68 matches played, 52 of which were correctly predicted, a success rate of 76.5%.
Here are the predictions for last week’s games.

Predictions for Round 8 Delayed Match

Here are the predictions for Round 8 Delayed Match. The prediction is my estimated expected points difference with a positive margin being a win to the home team, and a negative margin a win to the away team.

• Self-selected responses: this Twitter thread by Patrick Tomlinson is about the large number of negative reviews his book has on Goodreads. The book is still being edited. It won’t be available for months.
• From the Washington Post: “Our privacy experiment found that automakers collect data through hundreds of sensors and an always-on Internet connection. Driving surveillance is becoming hard to avoid”.
• Via the Herald: surprising no-one, a study by the US National Institute of Standards and Technology has found that US facial-recognition algorithms are better at recognising white people.  (Algorithms developed in east Asia seem to be ok at east Asian faces.)
• Last year I mentioned investigations at newsroom.co.nz by Eloise Gibson and others of Sir Ray Avery. Sir Ray filed a Media Council complaint about a story that alleged he had threatened a researcher with legal action. The Media Council has found in favour of newsroom
• From Radio NZ “Plans to collect data by putting sensors in thousands of state houses could result in the information being used to cut benefit payments or even evict tenants, a charity familiar with the project says.” The goal of the sensors is to find out what the actual heating/ventilation/damp problems are with state houses, information that would generalise to many other NZ houses. That’s a valuable goal. And Kāinga Ora say they don’t want to do anything else with the data, such as identify overcrowding or check who regularly has visitors staying the night,  or other potentially creepy possibilities.  But there doesn’t seem to be any clear mechanism to stop Kāinga Ora changing their minds.   Given the potential public benefits from suitable analyses of the data, it would be worth setting up a mechanism that people could see was trustworthy, rather than ending up with crap data when too many people opt out.
• Interpreting survey responses: The New York Times had an interactive clicky thing asking people to identify celebrities from their photographs.  Pete Buttigieg’s name was spelt in 268 different ways (click to embiggen). Presumably these weren’t all serious, but that doesn’t actually make life any easier for the survey analyst.
  

Auckland went impressively dark just before 2pm this afternoon: from Christina Hood on Twitter:

The reason is smoke from the Australian bushfires.

Fortunately for Auckland, the smoke (at the moment) is mostly at high altitude; you can tell because the lower picture is about as clear as the upper picture, just darker and more orange. Fine air pollution particles scatter light very effectively — in places with less humid summers than Auckland, light scattering is a good way to measure how much there is.

We can also look at the hourly measurements of PM2.5; fine-particle air pollution, or, basically, smoke.  Here are data from Penrose, in south central Auckland.

and Patumahoe, in far south Auckland

I chose these two because they were available and because there won’t be that much NZ-origin air pollution at those sites (on a Sunday afternoon).  At both locations there’s been an increase in fine particle air pollution, but not to any level of health concern.  It might be worth looking later in the evening, if you’re worried.  Given the placebo effect, if you don’t have specific health concerns you might be better off not checking.

(The Environment Auckland permalinks don’t work, so you need to go here and then to the AQ locations link, and then drill down.  You want “PM2.5 Hourly Aggregate unverified”)

This smoke used to be trees, 2000km away. We’ve seen particles from further than that — the Puyehue-Cordon Caulle volcano in Chile disrupted flights in 2011 — but the density of this smoke is much, much higher. WeatherWatch doesn’t know of a precedent.