Posts from February 2019 (17)

Team Ratings for Round 3

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 14 matches played, 10 of which were correctly predicted, a success rate of 71.4%.
Here are the predictions for last week’s games.

Predictions for Round 3

Here are the predictions for Round 3. 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 15

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 84 matches played, 60 of which were correctly predicted, a success rate of 71.4%.
Here are the predictions for last week’s games.

Predictions for Round 15

Here are the predictions for Round 15. 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 17

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 112 matches played, 88 of which were correctly predicted, a success rate of 78.6%.
Here are the predictions for last week’s games.

Predictions for Round 17

Here are the predictions for Round 17. 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.

Back in June, the AA put out a press release claiming that road deaths caused by (illegal) drugs now exceeded those caused by alcohol.  As I wrote at the time, (a) this was largely an artefact of increased testing, and (b) it’s a dodgy definition of ’cause’.

Roger Brooking, a drug and alcohol counsellor, has recently been pointing out that the definition of ’cause’ wasn’t comparable between alcohol and illegal drugs.  He’s right. Unfortunately, I think he wants the definitions changed in the wrong way.

The definition of ’caused by alcohol’ was ‘blood alcohol over the legal limit’, and the definition for illegal drugs was ‘detectable amount’.  There’s a sense in which these are comparable — they both indicate that violations of the law have occurred — but that’s not a useful sense if we’re talking about attributing risk.

Roger Brooking’s suggestion is to set both definitions at ‘any detectable’, and to set the legal alcohol limit for drivers to zero, and to do a lot more testing. I’m one of the minority of New Zealanders who wouldn’t be directly affected by any of these changes, but I still think it’s a terrible idea.

Focusing on statistical questions, though, describing as ’caused by’ drugs or alcohol all road deaths where the driver had detectable amounts in their blood is unambiguously wrong.  If you have alcohol in your system, your risk of being in a crash increases. It increases a moderate amount for moderate levels of alcohol, and by a lot for high levels of alcohol.  At very low levels the direct evidence is limited, but your risk probably increases a little bit. And with no alcohol or drugs in your system, the risk isn’t zero.  We know that from common sense, but also because some drivers do have crashes and tests don’t show any alcohol or drugs.

So, how much? The most recent research cited by a WHO report (PDF) is from the US, carried out in the late 1990s (PDF). The researchers monitored crashes in two US cities, and a week after each crash picked two random drivers at the time and place of the crash to measure their blood alcohol.  After adjusting for differences in, eg, age and adjusting for refusal to participate, they have estimates of the increase in risk for a range of blood alcohol concentrations. It’s worth noting that these are estimates rather than definitive truth, and that they are higher than previous estimates.    At 0.03% blood alcohol the estimated risk is 1.06 times higher; at 0.05% it’s 1.38 times higher; at 0.08% it’s 2.69 times higher; at 0.10% it’s 3.79 times higher.

That is, at 0.03% blood alcohol the estimate is 100 crashes that would have happened with no alcohol for every 6 influenced by alcohol. At 0.05% it’s 100 crashes that would have happened with no alcohol for every 38 influenced by alcohol. At 0.08%, it’s 100 crashes that would have happened with no alcohol for every 169 influenced by alcohol.  By the time you get to 0.10%, about three-quarters of the crashes are caused by alcohol, and at 0.15%, with a relative rate of 22 it’s damn near all of them. But at the legal limit it’s about a third, and at lower blood alcohol concentrations it’s a small fraction. If there were 80 fatal crashes where the driver had a blood alcohol above zero but below the legal limit, most of those crashes were not caused by alcohol in either a usual or a technical meaning of the word ’cause’.

Given the way risk decreases at lower levels of exposure to drugs, how should the numbers be reported? It makes sense to report both alcohol >0 and alcohol > legal limit. If there were consensus on the excess risk relationship it would be helpful to report crashes attributable to alcohol, taking the risk relationship into account.  It’s hard to do that for other drugs, though.  We’ve got very little empirical data relating forensic measures to risk for other drugs, and in some cases (eg cannabis) we know that the easily measurable concentration doesn’t pick up impairment well.  On top of that, use of multiple drugs is probably a substantial component of the problem.   It’s a good idea to increase measurements after serious crashes (to gather data). It’s a good idea to report alcohol + other drugs separately from alcohol alone and other drugs alone. And in the absence of any better criteria it’s probably unavoidable to just report any detectable level of other drugs, but we should resist (as ESR, notably, does) the temptation to call those crashes ’caused by’ the drug.

From RadioNZ this morning, Ben Thomas on the latest polling results

I think there’s a bit of a caution. Both of these polls came very shortly after the summer break, and when you look at polls over a year, the government of the day does best when people feel best about themselves. When do people feel best about themselves? Well, it’s when they’re on holiday, when they’re looking at barbeques… when the sun is shining.

That’s certainly reasonable. You could also think of other regions summer might be different, too. For example, with schools on holiday, you might get a different range of people being at home and answering the phone. In any case I wanted to see how much it shows up in the published opinion polls.

Peter Ellis has collected polling data from September 2002 right through to the last election, so I used that. Now, popularity of the government of the day has varied over this time, so I subtracted off a party difference, differences between polling companies, and a long-term time trend. Here’s the left-over variation when the long-term trend was averaged over about 5 years

and here’s when the long-term trend was over more like one year

The purple line estimates the seasonal variation, from a linear regression model.  Polls do seem more favorable to the incumbent during summer, but it’s a very small effect. The summer to winter difference is about half a percentage point, and there’s only fairly weak evidence that it’s in that direction rather than some other direction.

Here’s the same information, but wrapped around a yearly circle, with the points coloured according to whether Labour or National was in government at the time.  The black line is a circle corresponding to zero on the graphs above; the purple line shows the seasonal difference. If you look closely, you can see the purple sticks out to the sides more than the black: summer is more positive, winter is more negative. The labels are at January 1 for summer and then regularly spaced through the year.

What you do see clearly in this format is that people don’t do many polls around the new year. But the seasonal difference in results (for party intention, in publicly-released opinion polls) seems pretty small.

Team Ratings for Round 2

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 7 matches played, 5 of which were correctly predicted, a success rate of 71.4%.
Here are the predictions for last week’s games.

Predictions for Round 2

Here are the predictions for Round 2. 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 14

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 78 matches played, 57 of which were correctly predicted, a success rate of 73.1%.
Here are the predictions for last week’s games.

Predictions for Round 14

Here are the predictions for Round 14. 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 16

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 105 matches played, 82 of which were correctly predicted, a success rate of 78.1%.
Here are the predictions for last week’s games.

Predictions for Round 16

Here are the predictions for Round 16. 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.

From the Herald today:

That last number doesn’t look right.  At the 2013 Census, there were just under 90,000 people ordinarily resident in NZ who were born in the People’s Republic of China. Since then, there have been a net 46000 permanent or long-term migrants, according to a Stats NZ app — and recent research from Stats NZ has found that these figures overstate net migration a bit, because they misclassify some people returning home.  So, there are maybe 135,000 people living in NZ who were born in the PRC. Not all of these will think of NZ as home — some of them will be just here to study, for example — but it’s a reasonable group to consider. It’s not 290,000, and I don’t see how you can get that number.

On Twitter this morning, Tze Ming Mok speculated that the number might be people of Chinese ethnicity, but as she said, even that is hard to get as high as 290,000. And, very importantly, other people of Chinese ethnicity don’t necessarily have favourable views of the PRC — though they (and other people of East and Southeast Asian ethnicity) do get the spillover from both anti-PRC sentiment and traditional racism.

• “Often these studies are not found out to be inaccurate until there’s another real big dataset that someone applies these techniques to and says ‘oh my goodness, the results of these two studies don’t overlap‘,” she said. Genevra Allen (who gave one of the inaugural Ihaka Lectures here in Auckland) on machine learning in science.
• Good piece by Jenny Nicholls from North and South on algorithm risks (based around a new book, Hello World, by Hannah Fry)
• From the open AI blog, about a new neural network algorithm for generating realistic text: “Due to concerns about large language models being used to generate deceptive, biased, or abusive language at scale, we are only releasing a much smaller version of GPT-2 along with sampling code. We are not releasing the dataset, training code, or GPT-2 model weights.” Exercise for the reader: does this feel like a good idea? Would it feel like a good idea if Facebook were saying it?
• “PredPol claims to use an algorithm to predict crime in specific 500-foot by 500-foot sections of a city, so that police can patrol or surveil specific areas more heavily.” And they say that when police go to these areas they really do find crimes occurring there. Which…is less reassuring than PredPol seems to think.
• ” For example, a tench (a very big fish) is typically recognized by fingers on top of a greenish background. Why? Because most images in this category feature a fisherman holding up the tench like a trophy. “ About how neural networks work (a bit technical).
• Julian Sanchez argues that, yes, online click-through agreements are bad for data privacy, but partly because data consent is genuinely a hard problem