Thursday, 16 March 2017

Trust me, I'm a doctor...

…but don’t trust me if I’m an economist - that’s the clear message from a YouGov poll last month:

Note that the question asked is not just “do you trust them”, it’s “do you trust them when they talk about their field of expertise”. That the score is so low for economists is pretty damning.

I grabbed the YouGov data to have a look at what groups were driving the lack of trust in economists. It’s bad news:

Is it too much to expect the public to trust a group of experts on as complicated and nuanced a subject as economics? Science has one big similarity to economics which is relevant here: it calls on its practitioners to communicate the results of complicated models to a non-specialised audience and to policymakers. Science has its own share of contentious issues (at least they’re contentious to some): nuclear power, genetically modified food, vaccines, and climate change. So do people trust scientists? The short answer is overwhelmingly yes:

This is far from a perfect comparison. One of the arguments I’ve heard about the difference is that most non-experts would not find fault with what an expert said on, say, supersymmetry, but would find fault when the topic relates to their daily life, particularly if it is contrary to that individual’s personal experience. Perhaps there’s something in that. A bigger problem is that economics is a lot more politicised than science in general, and involvement with politics seems to be a good indicator for a poor net trust score in the first plot. There are also fewer certainties in economics, which muddies the waters. All that said, economics as a profession could undoubtedly do better and make up some of the gap. I think emulating some of the tactics of science communication could go some way in doing this.

Thursday, 2 March 2017

The social influence of genes

It's no surprise that beliefs and behaviours shared within social groups mean that one individual's preferences can influence outcomes for another individual. The Ising model, appropriated from physics, has been deployed to capture economic herding due to social influence (a short review may be found here (£) or here on the arXiv), while network theory has been used to get at the wider system of bilateral relationships between individuals. Schelling's model of segregation, though simple, shows how local social preferences can lead to effects which manifest at the macro level. Social effects are important. But what drives social influence? The need to 'fit-in'? That's surely part of it. But it may also have an underlying genetic component.

A fascinating recent paper in PLoS One measures the social influence due to genetics. They showed that the specific genetic makeup of one mouse has an effect on the outcomes (for instance, wound healing rate) of another mouse sharing the same cage. They call these phenomena social genetic effects (SGE). From the abstract,

"We find that genetic variation in cage mates (i.e. SGE) explains up to 29% of the variation in anxiety, wound healing, immune function, and body weight. Hence our study uncovers an unexpectedly large influence of the social environment. Additionally, we show that ignoring SGE can severely bias estimates of direct genetic effects (effects of an individual's genotypes on its own phenotype), which has important implications for the study of the genetic basis of complex traits."

29% is an enormous effect! Would the same effects ever apply in people? You can imagine situations where they would, at least qualitatively. Anxiety is easily spread between people: for example, phobias and fears can be acquired through observation as well as through direct experience, as in the famous study where monkeys learned to fear snakes from videos of other monkeys being afraid of snakes. Immune function is heavily dependent on sleep, and sleep patterns are heavily dependent on genetics: if one person has genes which make them more of an 'owl', a 'lark' partner might have their sleep disrupted. It'll be interesting to see if follow up studies can tease out the size of SGE in people. There may be some circumstances, eg in health economics, where SGE are an important source of unobserved heterogeneity.

Monday, 12 December 2016

Serious trust issues?

PhDcomics.com has produced a plot to show the average number of authors per paper in each discipline. Actually Jorge, the creator, only used the latest ten papers in the top 5 journals in each field as determined by their H-index, and there are all kinds of reasons why that might not be representative. Despite that, the plot pretty much tells the story you would expect: literature on one end, and 'big science', involving decades of work and huge experiments, on the other. Economics is closer to literature, in the part labelled 'serious trust issues'.

There may be some natural reasons why the number of authors per paper is low in economics journals. However, it seems likely to me that the main reason is that the alphabetical ordering of authors forms an artificial barrier to collaboration between larger groups. It's also well known that this system creates alphabetical winners and losers. The approach in the natural sciences (the scientific method, if you like) is, in my view, better. In that case, there are two important positions: first and last. The primary author is first (eg the PhD student or junior researcher who actually did the work), and the supervisor (eg the person who brought in the funding or managed the research group) is last. Everyone else is in the middle according to descending size of contribution. This carries more useful information for the reader than the pseudo-randomness of alphabetical ordering and it doesn't produce so many perverse incentives for the authors and potential authors. The output seems likely to be better and more collaborative research.

But don't take my word for it (especially as my surname begins with a T): see the paper by Engels et al., who study this theoretically (just one of many papers looking at these effects, this one ironically being in alphabetical order). They conclude that alphabetical ordering produces "research of lower quality than is optimal and than would be achieved if co-authors were forced to use name ordering to signal relative contribution."

Perhaps economists should all trust each other more and order by relative contribution in journals?

Wednesday, 16 November 2016

What if machines could collude on price?

It's an interesting thought and one which, a few years ago, would have been more at home in science fiction. Now, with so much online pricing being determined by algorithms, it deserves to be taken seriously.

And indeed the issues are being taken seriously in a new book, Virtual Competition. I haven't got my hands on a copy, but there's a brief and informative write-up in the journal Sciencehttp://science.sciencemag.org/content/354/6312/560.full

The article points out that it could be very difficult for regulators to discern whether a 'black box' machine learning algorithm like a neural network (which could take all prices in the market into account) was colluding with other pricing algorithms or not. Another fascinating idea is that consumers who are more budget constrained (and have browser histories filled with price comparison websites) may be offered lower prices by the machines.

One idea related to this is that by changing the prices that similar online consumers see, you could get a pretty good idea of the elasticity of demand for individual products by consumer type. The lack of menu costs for online products, the ability to record how many people looked but did not buy the product, and the potentially vast number of consumers makes this a much more feasible experiment than in a physical retail environment. This isn't an original idea - a recent paper has done this for Uber: http://www.nber.org/papers/w22627 and has plotted out the demand curve. But it would be entirely possible to do it for almost all online products sold in large volumes.

I'd be interested to know of any other studies along similar lines curves.

Thursday, 13 October 2016

Machine learning on the tube

Another week, another Nature paper by DeepMind/Google:

This time they’ve had a crack at ‘symbolic reasoning’ using a ‘differentiable neural computer’ which seems to be a deep neural network with a working memory so that it can distinguish between general rules (to which the learning applies) and specific instances (which are stored in memory).
For example, this can answer questions like “Where is the football?” with “playground” given the sentence “John is in the playground. John picked up the football.” trained only on a database of actor, action, object and location phrases. As I understand it, the exact words don’t matter, only the structure of the sentence, because of the way the symbolic learning works.

Another application they showed was to create random London tube maps (stations and lines of the real tube network but randomly connected), train the machine on them, and then ask the machine shortest path questions for the real tube network.

Perhaps the most exciting part of the paper is the last section though; they promise to eventually share the code.

Sunday, 11 September 2016

Data storage goes viral

...well, bacterial at least. There's a nice article in Nature about the stats for storage media, promising that storing data in bacterial DNA could be faster and more efficient than current storage technologies (infographic below):

Hopefully this will allow the $ per megabyte price of data storage to drop even further (see below). Interestingly, this pattern of exponential price fall is quite common for mass produced technologies such as solar cells and hard drives, as described in this paper:

I look forward to seeing my data come alive.