## Archive for statistics

Posted in Bad Statistics, Cute Problems with tags , , , , on November 25, 2016 by telescoper

I just came across this interesting little problem recently and thought I’d share it here. It’s usually called the ‘Neyman-Scott’ paradox. Before going on it’s worth mentioning that Elizabeth Scott (the second half of Neyman-Scott) was an astronomer by background. Her co-author was Jerzy Neyman. As has been the case for many astronomers, she contributed greatly to the development of the field of statistics. Anyway, I think this example provides another good illustration of the superiority of Bayesian methods for estimating parameters, but I’ll let you make your own mind up about what’s going on.

The problem is fairly technical so I’ve done done a quick version in latex that you can download

here, but I’ve also copied into this post so you can read it below:

I look forward to receiving Frequentist Flak or Bayesian Benevolence through the comments box below!

## What does “Big Data” mean to you?

Posted in The Universe and Stuff with tags , , , , on April 7, 2016 by telescoper

On several occasions recently I’ve had to talk about Big Data for one reason or another. I’m always at a disadvantage when I do that because I really dislike the term.Clearly I’m not the only one who feels this way:

For one thing the term “Big Data” seems to me like describing the Ocean as “Big Water”. For another it’s not really just the how big the data set is that matters. Size isn’t everything, after all. There is much truth in Stalin’s comment that “Quantity has a quality all its own” in that very large data sets allow you to do things you wouldn’t even try with smaller ones, but it can be complexity rather than sheer size that also requires new methods of analysis.

The biggest event in my own field of cosmology in the last few years has been the Planck mission. The data set is indeed huge: the above map of the temperature pattern in the cosmic microwave background has no fewer than 167 million pixels. That certainly caused some headaches in the analysis pipeline, but I think I would argue that this wasn’t really a Big Data project. I don’t mean that to be insulting to anyone, just that the main analysis of the Planck data was aimed at doing something very similar to what had been done (by WMAP), i.e. extracting the power spectrum of temperature fluctuations:

It’s a wonderful result of course that extends the measurements that WMAP made up to much higher frequencies, but Planck’s goals were phrased in similar terms to those of WMAP – to pin down the parameters of the standard model to as high accuracy as possible. For me, a real “Big Data” approach to cosmic microwave background studies would involve doing something that couldn’t have been done at all with a smaller data set. An example that springs to mind is looking for indications of effects beyond the standard model.

Moreover what passes for Big Data in some fields would be just called “data” in others. For example, the Atlas Detector on the  Large Hadron Collider  represents about 150 million sensors delivering data 40 million times per second. There are about 600 million collisions per second, out of which perhaps one hundred per second are useful. The issue here is then one of dealing with an enormous rate of data in such a way as to be able to discard most of it very quickly. The same will be true of the Square Kilometre Array which will acquire exabytes of data every day out of which perhaps one petabyte will need to be stored. Both these projects involve data sets much bigger and more difficult to handle that what might pass for Big Data in other arenas.

Books you can buy at airports about Big Data generally list the following four or five characteristics:

1. Volume
2. Velocity
3. Variety
4. Veracity
5. Variability

The first two are about the size and acquisition rate of the data mentioned above but the others are more about qualitatively different matters. For example, in cosmology nowadays we have to deal with data sets which are indeed quite large, but also very different in form.  We need to be able to do efficient joint analyses of hetergoeneous data structures with very different sampling properties and systematic errors in such a way that we get the best science results we can. Now that’s a Big Data challenge!

## The Insignificance of ORB

Posted in Bad Statistics with tags , , , on April 5, 2016 by telescoper

A piece about opinion polls ahead of the EU Referendum which appeared in today’s Daily Torygraph has spurred me on to make a quick contribution to my bad statistics folder.

The piece concerned includes the following statement:

David Cameron’s campaign to warn voters about the dangers of leaving the European Union is beginning to win the argument ahead of the referendum, a new Telegraph poll has found.

The exclusive poll found that the “Remain” campaign now has a narrow lead after trailing last month, in a sign that Downing Street’s tactic – which has been described as “Project Fear” by its critics – is working.

The piece goes on to explain

The poll finds that 51 per cent of voters now support Remain – an increase of 4 per cent from last month. Leave’s support has decreased five points to 44 per cent.

This conclusion is based on the results of a survey by ORB in which the number of participants was 800. Yes, eight hundred.

How much can we trust this result on statistical grounds?

Suppose the fraction of the population having the intention to vote in a particular way in the EU referendum is $p$. For a sample of size $n$ with $x$ respondents indicating that they hen one can straightforwardly estimate $p \simeq x/n$. So far so good, as long as there is no bias induced by the form of the question asked nor in the selection of the sample which, given the fact that such polls have been all over the place seems rather unlikely.

A little bit of mathematics involving the binomial distribution yields an answer for the uncertainty in this estimate of $p$ in terms of the sampling error:

$\sigma = \sqrt{\frac{p(1-p)}{n}}$

For the sample size of 800 given, and an actual value $p \simeq 0.5$ this amounts to a standard error of about 2%. About 95% of samples drawn from a population in which the true fraction is $p$ will yield an estimate within $p \pm 2\sigma$, i.e. within about 4% of the true figure. In other words the typical variation between two samples drawn from the same underlying population is about 4%. In other other words, the change reported between the two ORB polls mentioned above can be entirely explained by sampling variation and does not at all imply any systematic change of public opinion between the two surveys.

I need hardly point out that in a two-horse race (between “Remain” and “Leave”) an increase of 4% in the Remain vote corresponds to a decrease in the Leave vote by the same 4% so a 50-50 population vote can easily generate a margin as large as  54-46 in such a small sample.

Why do pollsters bother with such tiny samples? With such a large margin error they are basically meaningless.

I object to the characterization of the Remain campaign as “Project Fear” in any case. I think it’s entirely sensible to point out the serious risks that an exit from the European Union would generate for the UK in loss of trade, science funding, financial instability, and indeed the near-inevitable secession of Scotland. But in any case this poll doesn’t indicate that anything is succeeding in changing anything other than statistical noise.

Statistical illiteracy is as widespread amongst politicians as it is amongst journalists, but the fact that silly reports like this are commonplace doesn’t make them any less annoying. After all, the idea of sampling uncertainty isn’t all that difficult to understand. Is it?

And with so many more important things going on in the world that deserve better press coverage than they are getting, why does a “quality” newspaper waste its valuable column inches on this sort of twaddle?

## The Essence of Cosmology is Statistics

Posted in The Universe and Stuff with tags , , on September 8, 2015 by telescoper

I’m grateful to Licia Verde for sending this picture of me in action at last week’s conference in Castiglioncello.

The quote is one I use quite regularly, as the source is quite surprising. It is by George McVittie and appears in the Preface to the Proceedings of the Third Bekeley Symposium on Mathematical Statistics and Probability, which took place in 1956. It is surprising for two reasons. One is that McVittie is more strongly associated with theoretical cosmology than with statistics. In fact I have one of his books, the first edition of which was published in 1937:

There’s a bit in the book about observational cosmology, but basically it’s wall-to-wall Christoffel symbols!

The other surprising thing is that way back in 1956 there was precious little statistical information relevant to cosmology anyway, a far cry from the situation today with our plethora of maps and galaxy surveys. What he was saying though was that statistics is all about making inferences based on partial or incomplete data. Given that the subject of cosmology is the entire Universe, it is obvious we will never have complete data (i.e. we will never know everything). Hence cosmology is essentially statistical. This is true of other fields too, but in cosmology it is taken to an extreme. George McVittie passed away in 1988, so didn’t really live long enough to see this statement fulfilled, but it certainly has been over the last couple of decades!

P.S. Although he spent much of his working life in the East End of London (at Queen Mary College), George McVittie should not be confused with the even more famous, or rather infamous, Jack McVitie.

## Adventures with the One-Point Distribution Function

Posted in Bad Statistics, Books, Talks and Reviews, Talks and Reviews, The Universe and Stuff with tags , , on September 1, 2015 by telescoper

As I promised a few people, here are the slides I used for my talk earlier today at the meeting I am attending. Actually I was given only 30 minutes and used up a lot of that time on two things that haven’t got much to do with the title. One was a quiz to identify the six famous astronomers (or physicists) who had made important contributions to statistics (Slide 2) and the other was on some issues that arose during the discussion session yesterday evening. I didn’t in the end talk much about the topic given in the title, which was about how, despite learning a huge amount about certain aspects of galaxy clustering, we are still far from a good understanding of the one-point distribution of density fluctuations. I guess I’ll get the chance to talk more about that in the near future!

P.S. I think the six famous faces should be easy to identify, so there are no prizes but please feel free to guess through the comments box!

## Statistics in Astronomy

Posted in Bad Statistics, The Universe and Stuff with tags , , , , , , , , , on August 29, 2015 by telescoper

A few people at the STFC Summer School for new PhD students in Cardiff last week asked if I could share the slides. I’ve given the Powerpoint presentation to the organizers so presumably they will make the presentation available, but I thought I’d include it here too. I’ve corrected a couple of glitches I introduced trying to do some last-minute hacking just before my talk!

As you will inferfrom the slides, I decided not to compress an entire course on statistical methods into a one-hour talk. Instead I tried to focus on basic principles, primarily to get across the importance of Bayesian methods for tackling the usual tasks of hypothesis testing and parameter estimation. The Bayesian framework offers the only mathematically consistent way of tackling such problems and should therefore be the preferred method of using data to test theories. Of course if you have data but no theory or a theory but no data, any method is going to struggle. And if you have neither data nor theory you’d be better off getting one of the other before trying to do anything. Failing that, you could always go down the pub.

Rather than just leave it at that I thought I’d append some stuff  I’ve written about previously on this blog, many years ago, about the interesting historical connections between Astronomy and Statistics.

Once the basics of mathematical probability had been worked out, it became possible to think about applying probabilistic notions to problems in natural philosophy. Not surprisingly, many of these problems were of astronomical origin but, on the way, the astronomers that tackled them also derived some of the basic concepts of statistical theory and practice. Statistics wasn’t just something that astronomers took off the shelf and used; they made fundamental contributions to the development of the subject itself.

The modern subject we now know as physics really began in the 16th and 17th century, although at that time it was usually called Natural Philosophy. The greatest early work in theoretical physics was undoubtedly Newton’s great Principia, published in 1687, which presented his idea of universal gravitation which, together with his famous three laws of motion, enabled him to account for the orbits of the planets around the Sun. But majestic though Newton’s achievements undoubtedly were, I think it is fair to say that the originator of modern physics was Galileo Galilei.

Galileo wasn’t as much of a mathematical genius as Newton, but he was highly imaginative, versatile and (very much unlike Newton) had an outgoing personality. He was also an able musician, fine artist and talented writer: in other words a true Renaissance man.  His fame as a scientist largely depends on discoveries he made with the telescope. In particular, in 1610 he observed the four largest satellites of Jupiter, the phases of Venus and sunspots. He immediately leapt to the conclusion that not everything in the sky could be orbiting the Earth and openly promoted the Copernican view that the Sun was at the centre of the solar system with the planets orbiting around it. The Catholic Church was resistant to these ideas. He was hauled up in front of the Inquisition and placed under house arrest. He died in the year Newton was born (1642).

These aspects of Galileo’s life are probably familiar to most readers, but hidden away among scientific manuscripts and notebooks is an important first step towards a systematic method of statistical data analysis. Galileo performed numerous experiments, though he certainly didn’t carry out the one with which he is most commonly credited. He did establish that the speed at which bodies fall is independent of their weight, not by dropping things off the leaning tower of Pisa but by rolling balls down inclined slopes. In the course of his numerous forays into experimental physics Galileo realised that however careful he was taking measurements, the simplicity of the equipment available to him left him with quite large uncertainties in some of the results. He was able to estimate the accuracy of his measurements using repeated trials and sometimes ended up with a situation in which some measurements had larger estimated errors than others. This is a common occurrence in many kinds of experiment to this day.

Very often the problem we have in front of us is to measure two variables in an experiment, say X and Y. It doesn’t really matter what these two things are, except that X is assumed to be something one can control or measure easily and Y is whatever it is the experiment is supposed to yield information about. In order to establish whether there is a relationship between X and Y one can imagine a series of experiments where X is systematically varied and the resulting Y measured.  The pairs of (X,Y) values can then be plotted on a graph like the example shown in the Figure.

In this example on it certainly looks like there is a straight line linking Y and X, but with small deviations above and below the line caused by the errors in measurement of Y. This. You could quite easily take a ruler and draw a line of “best fit” by eye through these measurements. I spent many a tedious afternoon in the physics labs doing this sort of thing when I was at school. Ideally, though, what one wants is some procedure for fitting a mathematical function to a set of data automatically, without requiring any subjective intervention or artistic skill. Galileo found a way to do this. Imagine you have a set of pairs of measurements (xi,yi) to which you would like to fit a straight line of the form y=mx+c. One way to do it is to find the line that minimizes some measure of the spread of the measured values around the theoretical line. The way Galileo did this was to work out the sum of the differences between the measured yi and the predicted values mx+c at the measured values x=xi. He used the absolute difference |yi-(mxi+c)| so that the resulting optimal line would, roughly speaking, have as many of the measured points above it as below it. This general idea is now part of the standard practice of data analysis, and as far as I am aware, Galileo was the first scientist to grapple with the problem of dealing properly with experimental error.

The method used by Galileo was not quite the best way to crack the puzzle, but he had it almost right. It was again an astronomer who provided the missing piece and gave us essentially the same method used by statisticians (and astronomy) today.

Karl Friedrich Gauss (left) was undoubtedly one of the greatest mathematicians of all time, so it might be objected that he wasn’t really an astronomer. Nevertheless he was director of the Observatory at Göttingen for most of his working life and was a keen observer and experimentalist. In 1809, he developed Galileo’s ideas into the method of least-squares, which is still used today for curve fitting.

This approach involves basically the same procedure but involves minimizing the sum of [yi-(mxi+c)]2 rather than |yi-(mxi+c)|. This leads to a much more elegant mathematical treatment of the resulting deviations – the “residuals”.  Gauss also did fundamental work on the mathematical theory of errors in general. The normal distribution is often called the Gaussian curve in his honour.

After Galileo, the development of statistics as a means of data analysis in natural philosophy was dominated by astronomers. I can’t possibly go systematically through all the significant contributors, but I think it is worth devoting a paragraph or two to a few famous names.

I’ve already written on this blog about Jakob Bernoulli, whose famous book on probability was (probably) written during the 1690s. But Jakob was just one member of an extraordinary Swiss family that produced at least 11 important figures in the history of mathematics.  Among them was Daniel Bernoulli who was born in 1700.  Along with the other members of his famous family, he had interests that ranged from astronomy to zoology. He is perhaps most famous for his work on fluid flows which forms the basis of much of modern hydrodynamics, especially Bernouilli’s principle, which accounts for changes in pressure as a gas or liquid flows along a pipe of varying width.
But the elder Jakob’s work on gambling clearly also had some effect on Daniel, as in 1735 the younger Bernoulli published an exceptionally clever study involving the application of probability theory to astronomy. It had been known for centuries that the orbits of the planets are confined to the same part in the sky as seen from Earth, a narrow band called the Zodiac. This is because the Earth and the planets orbit in approximately the same plane around the Sun. The Sun’s path in the sky as the Earth revolves also follows the Zodiac. We now know that the flattened shape of the Solar System holds clues to the processes by which it formed from a rotating cloud of cosmic debris that formed a disk from which the planets eventually condensed, but this idea was not well established in the time of Daniel Bernouilli. He set himself the challenge of figuring out what the chance was that the planets were orbiting in the same plane simply by chance, rather than because some physical processes confined them to the plane of a protoplanetary disk. His conclusion? The odds against the inclinations of the planetary orbits being aligned by chance were, well, astronomical.

The next “famous” figure I want to mention is not at all as famous as he should be. John Michell was a Cambridge graduate in divinity who became a village rector near Leeds. His most important idea was the suggestion he made in 1783 that sufficiently massive stars could generate such a strong gravitational pull that light would be unable to escape from them.  These objects are now known as black holes (although the name was coined much later by John Archibald Wheeler). In the context of this story, however, he deserves recognition for his use of a statistical argument that the number of close pairs of stars seen in the sky could not arise by chance. He argued that they had to be physically associated, not fortuitous alignments. Michell is therefore credited with the discovery of double stars (or binaries), although compelling observational confirmation had to wait until William Herschel’s work of 1803.

It is impossible to overestimate the importance of the role played by Pierre Simon, Marquis de Laplace in the development of statistical theory. His book A Philosophical Essay on Probabilities, which began as an introduction to a much longer and more mathematical work, is probably the first time that a complete framework for the calculation and interpretation of probabilities ever appeared in print. First published in 1814, it is astonishingly modern in outlook.

Laplace began his scientific career as an assistant to Antoine Laurent Lavoiser, one of the founding fathers of chemistry. Laplace’s most important work was in astronomy, specifically in celestial mechanics, which involves explaining the motions of the heavenly bodies using the mathematical theory of dynamics. In 1796 he proposed the theory that the planets were formed from a rotating disk of gas and dust, which is in accord with the earlier assertion by Daniel Bernouilli that the planetary orbits could not be randomly oriented. In 1776 Laplace had also figured out a way of determining the average inclination of the planetary orbits.

A clutch of astronomers, including Laplace, also played important roles in the establishment of the Gaussian or normal distribution.  I have also mentioned Gauss’s own part in this story, but other famous astronomers played their part. The importance of the Gaussian distribution owes a great deal to a mathematical property called the Central Limit Theorem: the distribution of the sum of a large number of independent variables tends to have the Gaussian form. Laplace in 1810 proved a special case of this theorem, and Gauss himself also discussed it at length.

A general proof of the Central Limit Theorem was finally furnished in 1838 by another astronomer, Friedrich Wilhelm Bessel– best known to physicists for the functions named after him – who in the same year was also the first man to measure a star’s distance using the method of parallax. Finally, the name “normal” distribution was coined in 1850 by another astronomer, John Herschel, son of William Herschel.

I hope this gets the message across that the histories of statistics and astronomy are very much linked. Aspiring young astronomers are often dismayed when they enter research by the fact that they need to do a lot of statistical things. I’ve often complained that physics and astronomy education at universities usually includes almost nothing about statistics, because that is the one thing you can guarantee to use as a researcher in practically any branch of the subject.

Over the years, statistics has become regarded as slightly disreputable by many physicists, perhaps echoing Rutherford’s comment along the lines of “If your experiment needs statistics, you ought to have done a better experiment”. That’s a silly statement anyway because all experiments have some form of error that must be treated statistically, but it is particularly inapplicable to astronomy which is not experimental but observational. Astronomers need to do statistics, and we owe it to the memory of all the great scientists I mentioned above to do our statistics properly.