Posts from January 2017 (25)

The New Year is traditionally a time for short-term, one-off attempts to improve one’s health, like going to the gym for two weeks. One fashionable form is ‘detox’, where you take a few components of what might be a sensible change in diet, massively overdo them for a short time, then go back to your usual diet. The idea is that your body builds up ‘toxins’ that it’s unable to get rid of by normal biological processes, but that it can easily be tricked into getting rid of them rapidly by some special ritual.  Here’s a good piece from the Observer describing the problem. [update: and Michelle ‘Nanogirl’ Dickinson’s column this week, too]

The NZ media did ok on detox this year. There was a UK story about a particular herbal mixture causing dangerous sodium loss; there was one positive but somewhat restrained story; I’ve only seen one completely bogus one.

The moderately restrained story was in the Herald. It talked about a bunch of sensible dietary changes,  a bunch of basically unsupported herbal stuff, and for completeness, Native American sweat lodges. However, at least the main idea was to make long-term changes in one’s diet rather than to have some magical purification experience.  The story even had a couple of links to scientific papers, though they were to research showing that pollution might be harmful, which is not the problematic component of the detox myth.

On the other hand, on Twitter today, Peter Green posted a headline from the cover of “M2” magazine: “Six manly foods to detox your liver”. No, I’m not making this up.

It may help to know that other recent health headlines include “Experts Say Wearing This Colour Will Help You Have A More Effective Workout”, and “Neuroscience Says That This Song Reduces Anxiety By 65%”

If you’re wondering what detox foods are considered “manly” in the 21st century, the list includes turmeric, green tea, and broccoli sprouts. Quiche is still out.

Q: Do we need to move?

A: Um. No?

Q: “Living near a busy road could cause dementia”, it says.

A: No, that’s ok. We don’t live near a busy road, even to the extent that rhetorical constructs live anywhere.

Q: So what’s a ‘busy’ road, then?

A: An arterial or highway.  We’re more than 100m from the nearest one.

Q: That doesn’t sound very far.

A: The 1.07 times higher risk estimated in the research paper is for people within 50m of a major road.

Q: How many people is that?

A: In Ontario (mostly Toronto), 20%.  In Auckland, not so many.

Q: But we’re supposed to be in favour of population density and cities, aren’t we?

A: Yes. But even if the effect is real, it’s pretty small.

Q: The story says roads are responsible for 1 in 9 cases. That’s not so small.

A: One in 9 cases among people who live within 50m of a major road. Or, using one of the other estimates from the research, one in 14 cases among people who live within 50m of a major road.

Q: And 150m from a major road?

A: About one in 50 cases.

Q: Ok, that’s pretty small. Can they really detect it?

A: They’ve got data on a quarter of a million cases of dementia, so it’s borderline.

Q: But still?

A: Well, the the statistical evidence isn’t all that strong. A p-value of 0.035 from one of the three neurological diseases they looked at, isn’t much in a data set that large.

Q: And it’s just a correlation, right?

A: They’ve been able to do a reasonable job of removing other factors, and the road proximity was measured a long time before the dementia, so at least they don’t have cause and effect backwards.  But, yes, it could be something they didn’t have good enough data or modelling for.

Q: How about age? That’s a big issue with modelling dementia, isn’t it?

A: These are epidemiologists — “physicians broken down by age and sex”, as the old joke says — they know about age. They only compared groups of people of exactly the same age.

Q: But what does ‘exactly the same age’ even mean for something that doesn’t have a precise starting time?

A: That’s more of a problem. If people living near major roads got dementia at the same rate, but had it diagnosed six months earlier on average, that would be enough to explain the difference. There’s no particular reason that should happen, but it’s not impossible.

Q: So is the research worth looking at?

A: Worth looking at for consenting scientists in private, but not really worth international publicity.

Each week, we would like to invite readers of Stats Chat to submit nominations for our Stat of the Week competition and be in with the chance to win an iTunes voucher.

Here’s how it works:

• Anyone may add a comment on this post to nominate their Stat of the Week candidate before midday Friday January 6 2017.
• Statistics can be bad, exemplary or fascinating.
• The statistic must be in the NZ media during the period of December 31 2016 – January 6 2017 inclusive.
• Quote the statistic, when and where it was published and tell us why it should be our Stat of the Week.

Next Monday at midday we’ll announce the winner of this week’s Stat of the Week competition, and start a new one.

(more…)

If you’d like to comment on or debate any of this week’s Stat of the Week nominations, please do so below!

Machine learning/statistical learning has a very strong tendency to encode existing biases, because it works by finding patterns in existing data.  The ability to find patterns is very strong, and simply leaving out a variable you don’t want used isn’t enough if there are ways to extract the same information from other data. Because computers look objective and impartial, it can be easier to just accept their decisions — or regulations or trade-secret agreements may make it impossible to find out what they were doing.

That’s not necessarily a fatal flaw. People learn from existing cases, too. People can substitute a range of subtler social signals for crude, explicit bigotry.  It’s hard to force people to be honest about how they made a decision — they may not even know. Computer programs have the advantage of being much easier to audit for bias given the right regulatory framework; people have the advantage of occasionally losing some of their biases spontaneously.

Audit of black-box algorithms can be done in two complementary ways. You can give them made-up examples to see if differences that shouldn’t matter do affect the result, and you can see if their predictions on real examples were right.  The second is harder: if you give a loan to John from Epsom but not to Hone from Ōtara, you can see if John paid on time, but not if Hone would have.  Still, it can be done either using historical data or by just approving some loans that the algorithm doesn’t like.  You then need to decide whether the results were fair. That’s where things get surprisingly difficult.

Here’s a picture from a Google interactive

People are divided into orange and blue, with different distributions of credit scores. In this case the blue and orange people are equally likely on average to pay off a loan, but the credit score is more informative in orange people.  I’ve set the threshold so that the error rate of the prediction is the same in blue people as in orange people, which is obviously what you want. I could also have set the threshold so the proportion of approvals among people who would pay back the loan was the same in blue and orange people. That’s obviously what you want.  Or so the proportion of rejections among people who wouldn’t pay back the loan is the same. That, too, is obviously what you want.

You can’t have it all.

This isn’t one of the problems specific to social bias or computer algorithms or inaccurate credit scoring or evil and exploitative banks.  It’s a problem with any method of making decisions.  In fact, it’s a problem with any approach to comparing differences. You have to decide what summary of the difference you care about, because you can’t make them all the same.  This is old news in medical diagnostics, but appears not to have been considered in some other areas.

The motivation for my post was a post at Pro Publica on biases in automated sentencing decisions.  An earlier story had compared the specificity of the decisions according to race:  black people who didn’t end up reoffending were more likely to have been judged high risk than white people who didn’t end up reoffending. The company who makes the algorithm said, no, everything is fine because people who were judged high risk were equally likely to reoffend regardless of race. Both Pro Publica and the vendors are right on the maths; obviously they can’t both be right on the policy implications. We need to decide what we mean by a fair sentencing system. Personally, I’m not sure risk of reoffending should actually be a criterion, but if we stipulate that it is, there’s a decision to make.

In the new post, Julia Angwin and Jeff Larsen say

The findings were described in scholarly papers published or circulated over the past several months. Taken together, they represent the most far-reaching critique to date of the fairness of algorithms that seek to provide an objective measure of the likelihood a defendant will commit further crimes.

That’s true, but ‘algorithms’ and ‘objective’ don’t come into it. Any method of deciding who to release early has this problem, from judicial discretion in sentencing to parole boards to executive clemency. The only way around it is mandatory non-parole sentences, and even then you have to decide who gets charged with which crimes.

Fairness and transparency in machine learning are worth fighting for. They’re worth spending public money and political capital on. Part of the process must be deciding, with as much input from the affected groups as possible, what measures of fairness really matter to them. In the longer term, reducing the degree of disadvantage of, say, racial minorities should be the goal, and will automatically help with the decision problem. But a decision procedure that is ‘fair’ for disadvantaged groups both according to positive and negative predictive value and according to  specificity and sensitivity  isn’t worth fighting for, any more than a perpetual motion machine would be.