Bigger and Better (Data)?

Is bigger really better

I was browsing Data Science Central [1] recently and came across an article by Bill Vorhies, President & Chief Data Scientist of Data-Magnum. The piece was entitled 7 Cases Where Big Data Isn’t Better and is worth a read in full. Here I wanted to pick up on just a couple of Bill’s points.

In his preamble, he states:

Following the literature and the technology you would think there is universal agreement that more data means better models. […] However […] it’s always a good idea to step back and examine the premise. Is it universally true that our models will be more accurate if we use more data? As a data scientist you will want to question this assumption and not automatically reach for that brand new high-performance in-memory modeling array before examining some of these issues.

Bill goes on to make several pertinent points including: that if your data is bad, having more of it is not necessarily a solution; that attempting to create a gigantic and all-purpose model may well be inferior to multiple, more targeted models on smaller sub-sets of data; and that there exist specific instances where a smaller data sets yields greater accuracy [2]. However I wanted to pick up directly on Bill’s point 6 of 7, in which he also references Larry Greenemeier (@lggreenemeier) of Scientific American.

  Bill Vorhies   Larry Greenemeier  

6. Sometimes We Get Hypnotized By the Overwhelming Volume of the Data and Forget About Data Provenance and Good Project Design

A few months back I reviewed an article by Larry Greenemeier [3] about the failure of Google Flu Trend analysis to predict the timing and severity of flu outbreaks based on social media scraping. It was widely believed that this Big Data volume of data would accurately predict the incidence of flu but the study failed miserably missing timing and severity by a wide margin.

Says Greenemeier, “Big data hubris is the often the implicit assumption that big data are a substitute for, rather than a supplement to, traditional data collection and analysis. The mistake of many big data projects, the researchers note, is that they are not based on technology designed to produce valid and reliable data amenable for scientific analysis. The data comes from sources such as smartphones, search results and social networks rather than carefully vetted participants and scientific instruments”.

Perhaps more pertinent to a business environment, Greenemeier’s article also states:

Context is often lacking when info is pulled from disparate sources, leading to questionable conclusions.

Ruler

Neither of these authors is saying that having greater volumes of data is a definitively bad thing; indeed Vorhies states:

In general would I still prefer to have more data than less? Yes, of course.

They are however both pointing out that, in some instances, more traditional statistical methods, applied to smaller data sets yield superior results. This is particularly the case where data are repurposed and the use to which they are put is different to the considerations when they were collected; something which is arguably more likely to be the case where general purpose Big Data sets are leveraged without reference to other information.

Also, when large data sets are collated from many places, the data from each place can have different characteristics. If this variation is not controlled for in models, it may well lead to erroneous findings.

Statistical Methods

Their final observation is that sound statistical methodology needs to be applied to big data sets just as much as more regular ones. The hope that design flaws will simply evaporate when data sets get large enough may be seducing, but it is also dangerously wrong.

Vorhies and Greenemeier are not suggesting that Big Data has no value. However they state that one of its most potent uses may well be as a supplement to existing methods, perhaps extending them, or bringing greater granularity to results. I view such introspection in Data Science circles as positive, likely to lead to improved methods and an indication of growing maturity in the field. It is however worth noting that, in some cases, leverage of Small-but-Well-Designed Data [4] is not only effective, but actually a superior approach. This is certainly something that Data Scientists should bear in mind.
 


 
Notes

 
[1]
 
I’d recommend taking a look at this site regularly. There is a high volume of articles and the quality is variable, but often there are some stand-out pieces.
 
[2]
 
See the original article for the details.
 
[3]
 
The article was in Scientific American and entitled Why Big Data Isn’t Necessarily Better Data.
 
[4]
 
I may have to copyright this term and of course the very elegant abridgement, SBWDD.

 

 

How to be Surprisingly Popular

Popular with the Crowd
 
Introduction

This article is about the wisdom of the crowd [1], or more particularly its all too frequent foolishness. I am going to draw on a paper recently published in Nature by a cross-disciplinary team from the Massachusetts Institute of Technology and Princeton University. The authors are Dražen Prelec, H. Sebastian Seung and John McCoy. The paper’s title is A solution to the single-question crowd wisdom problem [2]. Rather than reinvent the wheel, here is a section from the abstract (with my emphasis):

Once considered provocative, the notion that the wisdom of the crowd is superior to any individual has become itself a piece of crowd wisdom, leading to speculation that online voting may soon put credentialed experts out of business. Recent applications include political and economic forecasting, evaluating nuclear safety, public policy, the quality of chemical probes, and possible responses to a restless volcano. Algorithms for extracting wisdom from the crowd are typically based on a democratic voting procedure. […] However, democratic methods have serious limitations. They are biased for shallow, lowest common denominator information, at the expense of novel or specialized knowledge that is not widely shared.

 
 
The Problems

The authors describe some compelling examples of where a crowd-based approach ignores the aforementioned specialised knowledge. I’ll cover a couple of these in a second, but let me first add my own.

How heavy is a proton?

Suppose we ask 1,000 people to come up with an estimate of the mass of a proton. One of these people happens to have won the Nobel Prize for Physics the previous year. Is the average of the estimates provided by the 1,000 people likely to be more accurate, or is the estimate of the one particularly qualified person going to be superior? There is an obvious answer to this question [3].

Lest it be thought that the above flaw in the wisdom of the crowd is confined to populations including a Nobel Laureate, I’ll reproduce a much more quotidian example from the Nature paper [4].

Philadelphia or Harrisburg?

[..] imagine that you have no knowledge of US geography and are confronted with questions such as: Philadelphia is the capital of Pennsylvania, yes or no? And, Columbia is the capital of South Carolina, yes or no? You pose them to many people, hoping that majority opinion will be correct. [in an actual exercise the team carried out] this works for the Columbia question, but most people endorse the incorrect answer (yes) for the Philadelphia question. Most respondents may only recall that Philadelphia is a large, historically significant city in Pennsylvania, and conclude that it is the capital. The minority who vote no probably possess an additional piece of evidence, that the capital is Harrisburg. A large panel will surely include such individuals. The failure of majority opinion cannot be blamed on an uninformed panel or flawed reasoning, but represents a defect in the voting method itself.

I’m both a good and bad example here. I know the capital of Pennsylvania is Harrisburg because I have specialist knowledge [5]. However my acquaintance with South Carolina is close to zero. I’d therefore get the first question right and have a 50 / 50 chance on the second (all other things being equal of course). My assumption is that Columbia is, in general, much more well-known than Harrisburg for some reason.

Confidence Levels

The authors go on to cover the technique that is often used to try to address this type of problem in surveys. Respondents are also asked how confident they are about their answer. Thus a tentative “yes” carries less weight than a definitive “yes”. However, as the authors point out, such an approach only works if correct responses are strongly correlated with respondent confidence. As is all too evident from real life, people are often both wrong and very confident about their opinion [6]. The authors extended their Philadelphia / Columbia study to apply confidence weightings, but with no discernible improvement.
 
 
A Surprisingly Popular Solution

As well as identifying the problem, the authors suggest a solution and later go on to demonstrate its efficacy. Again quoting from the paper’s abstract:

Here we propose the following alternative to a democratic vote: select the answer that is more popular than people predict. We show that this principle yields the best answer under reasonable assumptions about voter behaviour, while the standard ‘most popular’ or ‘most confident’ principles fail under exactly those same assumptions.

Let’s use the examples of capitals of states again here (as the authors do in the paper). As well as asking respondents, “Philadelphia is the capital of Pennsylvania, yes or no?” you also ask them “What percentage of people in this survey will answer ‘yes’ to this question?” The key is then to compare the actual survey answers with the predicted survey answers.

Columbia and Philadelphia [click to view a larger version in a new tab]

As shown in the above exhibit, in the authors’ study, when people were asked whether or not Columbia is the capital of South Carolina, those who replied “yes” felt that the majority of respondents would agree with them. Those who replied “no” symmetrically felt that the majority of people would also reply “no”. So no surprises there. Both groups felt that the crowd would agree with their response.

However, in the case of whether or not Philadelphia is the capital of Pennsylvania there is a difference. While those who replied “yes” also felt that the majority of people would agree with them, amongst those who replied “no”, there was a belief that the majority of people surveyed would reply “yes”. This is a surprise. People who make the correct response to this question feel that the wisdom of the crowd will be incorrect.

In the Columbia example, what people predict will be the percentage of people replying “yes” tracks with the actual response rate. In the Philadelphia example, what people predict will be the percentage of people replying “yes” is significantly less than the actual proportion of people making this response [7]. Thus a response of “no” to “Philadelphia is the capital of Pennsylvania, yes or no?” is surprisingly popular. The methodology that the authors advocate would then lead to the surprisingly popular answer (i.e. “no”) actually being correct; as indeed it is. Because there is no surprisingly popular answer in the Columbia example, then the result of a democratic vote stands; which is again correct.

To reiterate: a surprisingly popular response will overturn the democratic verdict, if there is no surprisingly popular response, the democratic verdict is unmodified.

Discriminating about Art

As well as confirming the superiority of the surprisingly popular approach (as opposed to either weighted or non-weighted democratic votes) with questions about state capitals, the authors went on to apply their new technique in a range of other areas [8].

  • Study 1 used 50 US state capitals questions, repeating the format [described above] with different populations [9].
     
  • Study 2 employed 80 general knowledge questions.
     
  • Study 3 asked professional dermatologists to diagnose 80 skin lesion images as benign or malignant.
     
  • Study 4 presented 90 20th century artworks [see the images above] to laypeople and art professionals, and asked them to predict the correct market price category.

Taking all responses across the four studies into account [10], the central findings were as follows [11]:

We first test pairwise accuracies of four algorithms: majority vote, surprisingly popular (SP), confidence-weighted vote, and max. confidence, which selects the answer endorsed with highest average confidence.

  • Across all items, the SP algorithm reduced errors by 21.3% relative to simple majority vote (P < 0.0005 by two-sided matched-pair sign test).
     
  • Across the items on which confidence was measured, the reduction was:
    • 35.8% relative to majority vote (P < 0.001),
    • 24.2% relative to confidence-weighted vote (P = 0.0107) and
    • 22.2% relative to max. confidence (P < 0.13).

The authors go on to further kick the tyres [12] on these results [13] without drawing any conclusions that deviate considerably from the ones they first present and which are reproduced above. The surprising finding is that the surprisingly popular algorithm significantly out-performs the algorithms normally used in wisdom of the crowd polling. This is a major result, in theory at least.
 
 
Some Thoughts

Tools and Toolbox

At the end of the abstract, the authors state that:

Like traditional voting, [the surprisingly popular algorithm] accepts unique problems, such as panel decisions about scientific or artistic merit, and legal or historical disputes. The potential application domain is thus broader than that covered by machine learning […].

Given the – justified – attention that has been given to machine learning in recent years, this is a particularly interesting claim. More broadly, SP seems to bring much needed nuance to the wisdom of the crowd. It recognises that the crowd may often be right, but also allows better informed minorities to override the crowd opinion in specific cases. It does this robustly in all of the studies that the authors conducted. It will be extremely interesting to see this novel algorithm deployed in anger, i.e. in a non-theoretical environment. If its undoubted promise is borne out – and the evidence to date suggests that it will be – then statisticians will have a new and powerful tool in their arsenal and a range of predictive activities will be improved.

The scope of applicability of the SP technique is as wide as that of any wisdom of the crowd approach and, to repeat the comments made by the authors in their abstract, has recently included:

[…] political and economic forecasting, evaluating nuclear safety, public policy, the quality of chemical probes, and possible responses to a restless volcano

If the author’s initial findings are repeated in “live” situations, then the refinement to the purely democratic approach that SP brings should elevate an already useful approach to being an indispensable one in many areas.

I will let the authors have a penultimate word [14]:

Although democratic methods of opinion aggregation have been influential and productive, they have underestimated collective intelligence in one respect. People are not limited to stating their actual beliefs; they can also reason about beliefs that would arise under hypothetical scenarios. Such knowledge can be exploited to recover truth even when traditional voting methods fail. If respondents have enough evidence to establish the correct answer, then the surprisingly popular principle will yield that answer; more generally, it will produce the best answer in light of available evidence. These claims are theoretical and do not guarantee success in practice, as actual respondents will fall short of ideal. However, it would be hard to trust a method [such as majority vote or confidence-weighted vote] if it fails with ideal respondents on simple problems like [the Philadelphia one]. To our knowledge, the method proposed here is the only one that passes this test.

US Presidential Election Polling [borrowed from Wikipedia]

The ultimate thought I will present in this article is an entirely speculative one. The authors posit that their method could be applied to “potentially controversial topics, such as political and environmental forecasts”, while cautioning that manipulation should be guarded against. Their suggestion leads me wonder what impact on the results of opinion polls a suitably formed surprisingly popular questionnaire would have had in the run up to both the recent UK European Union Referendum and the plebiscite for the US Presidency. Of course it is now impossible to tell, but maybe some polling organisations will begin to incorporate this new approach going forward. It can hardly make things worse.
 


 
Notes

 
[1]
 
According to Wikipedia, the phenomenon that:

A large group’s aggregated answers to questions involving quantity estimation, general world knowledge, and spatial reasoning has generally been found to be as good as, and often better than, the answer given by any of the individuals within the group.

The authors of the Nature paper question whether this is true in all circumstances.

 
[2]
 
Prelec, D., Seung, H.S., McCoy, J., (2017). A solution to the single-question crowd wisdom problem. Nature 541, 532–535.

You can view a full version of this paper care of Springer Nature SharedIt at the following link. ShareIt is Springer’s content sharing initiative.

Direct access to the article on Nature’s site (here) requires a subscription to the journal.

 
[3]
 
This example is perhaps an interesting rejoinder to the increasing lack of faith in experts in the general population, something I covered in Toast.

Of course the answer is approximately: 1.6726219 × 10-27 kg.

 
[4]
 
I have lightly edited this section but abjured the regular bracketed ellipses (more than one […] as opposed to conic sections as I note elsewhere). This is both for reasons of readability and also as I have not yet got to some points that the authors were making in this section. The original text is a click away.
 
[5]
 
My wife is from this state.
 
[6]
 
Indeed it sometimes seems that the more wrong the opinion, the more certain that people believe it to be right.

Here the reader is free to insert whatever political example fits best with their worldview.

 
[7]
 
Because many people replying “no” felt that a majority would disagree with them.
 
[8]
 
Again I have lightly edited this text.
 
[9]
 
To provide a bit of detail, here the team created a questionnaire with 50 separate questions sets of the type:

  1. {Most populous city in a state} is the capital of {state}: yes or no?
     
  2. How confident are you in your answer (50- 100%)?
     
  3. What percentage of people surveyed will respond “yes” to this question? (1 – 100%)

This was completed by 83 people split between groups of undergraduate and graduate students at both MIT and Princeton. Again see the paper for further details.

 
[10]
 
And eliding some nuances such as some responses being binary (yes/no) and others a range (e.g. the dermatologists were asked to rate the chance of malignancy on a six point scale from “absolutely uncertain to absolutely certain”). Also respondents were asked to provide their confidence in some studies and not others.
 
[11]
 
Once more with some light editing.
 
[12]
 
This is a technical term employed in scientific circles an I apologise if my use of jargon confuses some readers.
 
[13]
 
Again please see the actual paper for details.
 
[14]
 
Modified very slightly by my last piece of editing.

 

 

Toast

Acrylamide [borrowed from Wikipedia]

Foreword

This blog touches on a wide range of topics, including social media, cultural transformation, general technology and – last but not least – sporting analogies. However, its primary focus has always been on data and information-centric matters in a business context. Having said this, all but the more cursory of readers will have noted the prevalence of pieces with a Mathematical or Scientific bent. To some extent this is a simple reflection of the author’s interests and experience, but a stronger motivation is often to apply learnings from different fields to the business data arena. This article is probably more scientific in subject matter than most, but I will also look to highlight some points pertinent to commerce towards the end.
 
 
Introduction

In Science We Trust?

The topic I want to turn my attention to in this article is public trust in science. This is a subject that has consumed many column inches in recent years. One particular area of focus has been climate science, which, for fairly obvious political reasons, has come in for even more attention than other scientific disciplines of late. It would be distracting to get into the arguments about climate change and humanity’s role in it here [1] and in a sense this is just the latest in a long line of controversies that have somehow become attached to science. An obvious second example here is the misinformation circling around both the efficacy and side effects of vaccinations [2]. In both of these cases, it seems that at least a sizeable minority of people are willing to query well-supported scientific findings. In some ways, this is perhaps linked to the general mistrust of “experts” and “elites” [3] that was explicitly to the fore in the UK’s European Union Referendum debate [4].

“People in this country have had enough of experts”

– Michael Gove [5], at this point UK Justice Secretary and one of the main proponents of the Leave campaign, speaking on Sky News, June 2016.

Mr Gove was talking about economists who held a different point of view to his own. However, his statement has wider resonance and cannot be simply dismissed as the misleading sound-bite of an experienced politician seeking to press his own case. It does indeed appear that in many places around the world experts are trusted much less than they used to be and that includes scientists.

“Many political upheavals of recent years, such as the rise of populist parties in Europe, Donald Trump’s nomination for the American presidency and Britain’s vote to leave the EU, have been attributed to a revolt against existing elites.”

The Buttonwood column, The Economist, September 2016.

Why has this come to be?
 
 
A Brief [6] History of the Public Perception of Science

Public Perception

Note: This section is focussed on historical developments in the public’s trust in science. If the reader would like to skip on to more toast-centric content, then please click here.

Answering questions about the erosion of trust in politicians and the media is beyond the scope of this humble blog. Wondering what has happened to trust in science is firmly in its crosshairs. One part of the answer is that – for some time – scientists were held in too much esteem and the pendulum was inevitably going to swing back the other way. For a while the pace of scientific progress and the miracles of technology which this unleashed placed science on a pedestal from which there was only one direction of travel. During this period in which science was – in general – uncritically held in great regard, the messy reality of actual science was never really highlighted. The very phrase “scientific facts” is actually something of an oxymoron. What we have is instead scientific theories. Useful theories are consistent with existing observations and predict new phenomena. However – as I explained in Patterns patterns everywhere – a theory is only as good as the latest set of evidence and some cherished scientific theories have been shown to be inaccurate; either in general, or in some specific circumstances [7]. However saying “we have a good model that helps us explain many aspects of a phenomenon and predict more, but it doesn’t cover everything and there are some uncertainties” is a little more of a mouthful than “we have discovered that…”.

There have been some obvious landmarks along the way to science’s current predicament. The unprecedented destruction unleashed by the team working on the Manhattan Project at first made the scientists involved appear God-like. It also seemed to suggest that the path to Great Power status was through growing or acquiring the best Physicists. However, as the prolonged misery caused in Japan by the twin nuclear strikes became more apparent and as the Cold War led to generations living under the threat of mutually assured destruction, the standing attached by the general public to Physicists began to wane; the God-like mantle began to slip. While much of our modern world and its technology was created off the back of now fairly old theories like Quantum Chromodynamics and – most famously – Special and General Relativity, the actual science involved became less and less accessible to the man or woman in the street. For all the (entirely justified) furore about the detection of the Higgs Boson, few people would be able to explain much about what it is and how it fits into the Standard Model of particle physics.

In the area of medicine and pharmacology, the Thalidomide tragedy, where a drug prescribed to help pregnant women suffering from morning sickness instead led to terrible birth defects in more than 10,000 babies, may have led to more stringent clinical trials, but also punctured the air of certainty that had surrounded the development of the latest miracle drug. While medical science and related disciplines have vastly improved the health of much of the globe, the glacial progress in areas such as oncology has served as a reminder of the fallibility of some scientific endeavours. In a small way, the technical achievements of that apogee of engineering, NASA, were undermined by loss of crafts and astronauts. Most notably the Challenger and Columbia fatalities served to further remove the glossy veneer that science had acquired in the 1940s to 1960s.

Lest it be thought at this point that I am decrying science, or even being anti-scientific, nothing could be further from the truth. I firmly believe that the ever growing body of scientific knowledge is one of humankind’s greatest achievements, if not its greatest. From our unpromising vantage point on an unremarkable little planet in our equally common-all-garden galaxy we have been able to grasp many of the essential truths about the whole Universe from the incomprehensibly gigantic to the most infinitesimal constituent of a sub-atomic particle. However, it seems that many people do not fully embrace the grandeur of our achievements, or indeed in many cases the unexpected beauty and harmony that they have revealed [8]. It is to the task of understanding this viewpoint that I am addressing my thoughts.

More recently, the austerity that has enveloped much of the developed world since the 2008 Financial Crisis has had two reinforcing impacts on science in many countries. First funding has often been cut, leading to pressure on research programmes and scientists increasingly having to make an economic case for their activities; a far cry from the 1950s. Second, income has been effectively stagnant for the vast majority of people, this means that scientific expenditure can seem something of a luxury and also fuels the anti-elite feelings cited by The Economist earlier in this article.

Anita Makri

Into this seeming morass steps Anita Makri, “editor/writer/producer and former research scientist”. In a recent Nature article she argues that the form of science communicated in popular media leaves the public vulnerable to false certainty. I reproduce some of her comments here:

“Much of the science that the public knows about and admires imparts a sense of wonder and fun about the world, or answers big existential questions. It’s in the popularization of physics through the television programmes of physicist Brian Cox and in articles about new fossils and quirky animal behaviour on the websites of newspapers. It is sellable and familiar science: rooted in hypothesis testing, experiments and discovery.

Although this science has its place, it leaves the public […] with a different, outdated view to that of scientists of what constitutes science. People expect science to offer authoritative conclusions that correspond to the deterministic model. When there’s incomplete information, imperfect knowledge or changing advice — all part and parcel of science — its authority seems to be undermined. […] A popular conclusion of that shifting scientific ground is that experts don’t know what they’re talking about.”

– Anita Makri, Give the public the tools to trust scientists, Nature, January 2017.

I’ll come back to Anita’s article again later.
 
 
Food Safety – The Dangers Lurking in Toast

Food Safety

After my speculations about the reasons why science is held in less esteem than once was the case, I’ll return to more prosaic matters; namely food and specifically that humble staple of many a breakfast table, toast. Food science has often fared no better than its brother disciplines. The scientific guidance issued to people wanting to eat healthily can sometimes seem to gyrate wildly. For many years fat was the source of all evil, more recently sugar has become public enemy number one. Red wine was meant to have beneficial effects on heart health, then it was meant to be injurious; I’m not quite sure what the current advice consists of. As Makri states above, when advice changes as dramatically as it can do in food science, people must begin to wonder whether the scientists really know anything at all.

So where does toast fit in? Well the governmental body charged with providing advice about food in the UK is called the Food Standards Agency. They describe their job as “using our expertise and influence so that people can trust that the food they buy and eat is safe and honest.” While the FSA do sterling work in areas such as publicly providing ratings of food hygiene for restaurants and the like, their most recent campaign is one which seems at best ill-advised and at worst another nail in the public perception of the reliability of scientific advice. Such things matter because they contribute to the way that people view science in general. If scientific advice about food is seen as unsound, surely there must be questions around scientific advice about climate change, or vaccinations.

Before I am accused of belittling the FSA’s efforts, let’s consider the campaign in question, which is called Go for Gold and encourages people to consume less acrylamide. Here is some of what the FSA has to say about the matter:

“Today, the Food Standards Agency (FSA) is launching a campaign to ‘Go for Gold’, helping people understand how to minimise exposure to a possible carcinogen called acrylamide when cooking at home.

Acrylamide is a chemical that is created when many foods, particularly starchy foods like potatoes and bread, are cooked for long periods at high temperatures, such as when baking, frying, grilling, toasting and roasting. The scientific consensus is that acrylamide has the potential to cause cancer in humans.

[…]

as a general rule of thumb, aim for a golden yellow colour or lighter when frying, baking, toasting or roasting starchy foods like potatoes, root vegetables and bread.”

– Food Standards Agency, Families urged to ‘Go for Gold’ to reduce acrylamide consumption, January 2017.

The Go for Gold campaign was picked up by various media outlets in the UK. For example the BBC posted an article on its web-site which opened by saying:

Dangerous Toast [borrowed from the BBC]

“Bread, chips and potatoes should be cooked to a golden yellow colour, rather than brown, to reduce our intake of a chemical which could cause cancer, government food scientists are warning.”

– BBC, Browned toast and potatoes are ‘potential cancer risk’, say food scientists, January 2017.

The BBC has been obsessed with neutrality on all subjects recently [9], but in this case they did insert the reasonable counterpoint that:

“However, Cancer Research UK [10] said the link was not proven in humans.”

Acrylamide is certainly a nasty chemical. Amongst other things, it is used in polyacrylamide gel electrophoresis, a technique used in biochemistry. If biochemists mix and pour their own gels, they have to monitor their exposure and there are time-based and lifetime limits as to how often they can do such procedures [11]. Acrylamide has also been shown to lead to cancer in mice. So what could be more reasonable that the FSA’s advice?
 
 
Food Safety – A Statistical / Risk Based Approach

David Spiegelhalter

Earlier I introduced Anita Makri, it is time to meet our second protagonist, David Spiegelhalter, Winton Professor for the Public Understanding of Risk in the Statistical Laboratory, Centre for Mathematical Sciences, University of Cambridge [12]. Professor Spiegelhalter has penned a response to the FSA’s Go for Gold campaign. I feel that this merits reading in entirety, but here are some highlights:

“Very high doses [of Acrylamide] have been shown to increase the risk of mice getting cancer. The IARC (International Agency for Research on Cancer) considers it a ‘probable human carcinogen’, putting it in the same category as many chemicals, red meat, being a hairdresser and shift-work.

However, there is no good evidence of harm from humans consuming acrylamide in their diet: Cancer Research UK say that ‘At the moment, there is no strong evidence linking acrylamide and cancer.’

This is not for want of trying. A massive report from the European Food Standards Agency (EFSA) lists 16 studies and 36 publications, but concludes

  ‘In the epidemiological studies available to date, AA intake was not associated with an increased risk of most common cancers, including those of the GI or respiratory tract, breast, prostate and bladder. A few studies suggested an increased risk for renal cell, and endometrial (in particular in never-smokers) and ovarian cancer, but the evidence is limited and inconsistent. Moreover, one study suggested a lower survival in non-smoking women with breast cancer with a high pre-diagnostic exposure to AA but more studies are necessary to confirm this result. (p185)’
[…] [Based on the EFSA study] adults with the highest consumption of acrylamide could consume 160 times as much and still only be at a level that toxicologists think unlikely to cause increased tumours in mice.

[…]

This all seems rather reassuring, and may explain why it’s been so difficult to observe any effect of acrylamide in diet.”

– David Spiegelhalter, Opinion: How dangerous is burnt toast?, University of Cambridge, January 2017.

Indeed, Professor Spiegelhalter, an esteemed statistician, also points out that most studies will adopt the standard criteria for statistical significance. Given that such significance levels are often set at 5%, then this means that:

“[As] each study is testing an association with a long list of cancers […], we would expect 1 in 20 of these associations to be positive by chance alone.”

He closes his article by stating – not unreasonably – that the FSA’s time and attention might be better spent on areas where causality between an agent and morbidity is well-established, for example obesity. My assumption is that the FSA has a limited budget and has to pick and choose what food issues to weigh in on. Even if we accept for the moment that there is some slight chance of a causal link between the consumption of low levels of acrylamide and cancer, there are plenty of other areas in which causality is firmly established; obesity as mentioned by Professor Spiegelhalter, excessive use of alcohol, even basic kitchen hygiene. It is hard to understand why the FSA did not put more effort into these and instead focussed on an area where the balance of scientific judgement is that there is unlikely to be an issue.

Having a mathematical background perhaps biases me, but I tend to side with Professor Spiegelhalter’s point of view. I don’t want to lay the entire blame for the poor view that some people have of science at the FSA’s door, but I don’t think campaigns like Go for Gold help very much either. The apocryphal rational man or woman will probably deduce that there is not an epidemic of acrylamide poisoning in progress. This means that they may question what the experts at the FSA are going on about. In turn this reduces respect for other – perhaps more urgent – warnings about food and drink. Such a reaction is also likely to colour how the same rational person thinks about “expert” advice in general. All of this can contribute to further cracks appearing in the public edifice of science, an outcome I find very unfortunate.

So what is to be done?
 
 
A Call for a New and More Honest Approach to Science Communications

Honesty is the Best Policy

As promised I’ll return to Anita Makri’s thoughts in the same article referenced above:

“It’s more difficult to talk about science that’s inconclusive, ambivalent, incremental and even political — it requires a shift in thinking and it does carry risks. If not communicated carefully, the idea that scientists sometimes ‘don’t know’ can open the door to those who want to contest evidence.

[…]

Scientists can influence what’s being presented by articulating how this kind of science works when they talk to journalists, or when they advise on policy and communication projects. It’s difficult to do, because it challenges the position of science as a singular guide to decision making, and because it involves owning up to not having all of the answers all the time while still maintaining a sense of authority. But done carefully, transparency will help more than harm. It will aid the restoration of trust, and clarify the role of science as a guide.”

The scientific method is meant to be about honesty. You record what you see, not what you want to see. If the data don’t support your hypothesis, you discard or amend your hypothesis. The peer-review process is meant to hold scientists to the highest levels of integrity. What Makri seems to be suggesting is for scientists to turn their lenses on themselves and how they communicate their work. Being honest where there is doubt may be scary, but not as scary as being caught out pushing certainty where no certainty is currently to be had.
 


 
Epilogue

At the beginning of this article, I promised that I would bring things back to a business context. With lots of people with PhDs in numerate sciences now plying their trade as data scientists and the like, there is an attempt to make commerce more scientific [13]. Understandably, the average member of a company will have less of an appreciation of statistics and statistical methods than their data scientists do. This can lead to data science seeming like magic; the philosopher’s stone [14]. There are obvious parallels here with how Physicists were seen in the period immediately after the Second World War.

Earlier in the text, I mused about what factors may have led to a deterioration in how the public views science and scientists. I think that there is much to be learnt from the issues I have covered in this article. If data scientists begin to try to peddle absolute truth and perfect insight (both of which, it is fair to add, are often expected from them by non-experts), as opposed to ranges of outcomes and probabilities, then the same decline in reputation probably awaits them. Instead it would be better if data scientists heeded Anita Makri’s words and tried to always be honest about what they don’t know as well as what they do.
 


 
Notes

 
[1]
 
Save to note that there really is no argument in scientific circles.

As ever Randall Munroe makes the point pithily in his Earth Temperature Timeline – https://xkcd.com/1732/.

For a primer on the area, you could do worse than watching The Royal Society‘s video:

 
[2]
 
For the record, my daughter has had every vaccine known to the UK and US health systems and I’ve had a bunch of them recently as well.
 
[3]
 
Most scientists I know would be astonished that they are considered part of the amorphous, ill-defined and obviously malevolent global “elite”. Then “elite” is just one more proxy for “the other” something which it is not popular to be in various places in the world at present.
 
[4]
 
Or what passed for debate in these post-truth times.
 
[5]
 
Mr Gove studied English at Lady Margaret Hall, Oxford, where he was also President of the Oxford Union. Clearly Oxford produces less experts than it used to in previous eras.
 
[6]
 
One that is also probably wildly inaccurate and certainly incomplete.
 
[7]
 
So Newton’s celebrated theory of gravitation is “wrong” but actually works perfectly well in most circumstances. The the Rutherford–Bohr model, where atoms are little Solar Systems, with the nucleus circled by electrons much as the planets circle the Sun is “wrong”, but actually does serve to explain a number of things; if sadly not the orbital angular momentum of electrons.
 
[8]
 
Someone should really write a book about that – watch this space!
 
[9]
 
Not least in the aforementioned EU Referendum where it felt the need to follow the views of the vast majority of economists with those of the tiny minority, implying that the same weight be attached to both points of view. For example, 99.9999% of people believe the world to be round, but in the interests of balance my mate Jim reckons it is flat.
 
[10]
 
According to their web-site: “the world’s leading charity dedicated to beating cancer through research”.
 
[11]
 
As attested to personally by the only proper scientist in our family.
 
[12]
 
Unlike Oxford (according to Mr Gove anyway), Cambridge clearly still aspires to creating experts.
 
[13]
 
By this I mean proper science and not pseudo-science like management theory and the like.
 
[14]
 
In the original, non-J.K. Rowling sense of the phrase.

 

 

Indiana Jones and The Anomalies of Data

One of an occasional series [1] highlighting the genius of Randall Munroe. Randall is a prominent member of the international data community and apparently also writes some sort of web-comic as a side line [2].

I didn't even realize you could HAVE a data set made up entirely of outliers.
Copyright xkcd.com

Data and Indiana Jones, these are a few of my favourite things… [3] Indeed I must confess to having used a variant of the image below in each of my seminar deck and – on this site back in 2009 – a previous article, A more appropriate metaphor for Business Intelligence projects.

Raiders of the Lost Ark II would have been a much better title than Temple of Doom IMO

In both cases I was highlighting that data-centric work is sometimes more like archaeology than the frequently employed metaphor of construction. To paraphrase myself, you never know what you will find until you start digging. The image suggested the unfortunate results of not making this distinction when approaching data projects.

So, perhaps I am arguing for less Data Architects and more Data Archaeologists; the whip and fedora are optional of course!
 


 Notes

 
[1]
 
Well not that occasional as, to date, the list extends to:

  1. Patterns patterns everywhere – The Sequel
  2. An Inconvenient Truth
  3. Analogies, the whole article is effectively an homage to xkcd.com
  4. A single version of the truth?
  5. Especially for all Business Analytics professionals out there
  6. New Adventures in Wi-Fi – Track 1: Blogging
  7. Business logic [My adaptation]
  8. New Adventures in Wi-Fi – Track 2: Twitter
  9. Using historical data to justify BI investments – Part III
 
[2]
 
xkcd.com if you must ask.
 
[3]
 
Though in this case, my enjoyment would have been further enhanced by the use of “artefacts” instead.

 

 

More Statistics and Medicine

Weighing Medicine in the balance

I wrote last on the intersection of these two disciplines back in March 2011 (Medical Malpractice). What has prompted me to return to the subject is some medical tests that I was offered recently. If the reader will forgive me, I won’t go into the medical details – and indeed have also obfuscated some of the figures I was quoted – but neither are that relevant to the point that I wanted to make. This point relates to how statistics are sometimes presented in medical situations and – more pertinently – the disconnect between how these may be interpreted by the man or woman in the street, as opposed to what is actually going on.

Rather than tie myself in knots, let’s assume that the test is for a horrible disease called PJT Syndrome [1]. Let’s further assume that I am told that the test on offer has an accuracy of 80% [2]. This in and of itself is a potentially confusing figure. Does the test fail to detect the presence of PJT Syndrome 20% of the time, or does it instead erroneously detect PJT Syndrome, when the patient is actually perfectly healthy, 20% of the time? In this case, after an enquiry, I was told that a negative result was a negative result, but that a positive one did not always mean that the subject suffered from PJT Syndrome; so the issue is confined to false positives, not false negatives. This definition of 80% accuracy is at least a little clearer.

So what is a reasonable person to deduce from the 80% figure? Probably that if they test positive, that there is an 80% certainty that they have PJT Syndrome. I think that my visceral reaction would probably be along those lines. However, such a conclusion can be incorrect, particularly where the incidence of PJT Syndrome is low in a population. I’ll try to explain why.

If we know that PJT Syndrome occurs in 1 in every 100 people on average, what does this mean for the relevance of our test results? Let’s take a graphical look at a wholly representative population of exactly 100 people. The PJT Syndrome sufferer appears in red at the bottom right.

1 in 100

Now what is the result of the 80% accuracy of our test, remembering that this means that 20% of people taking it will be falsely diagnosed as having PJT Syndrome? Well 20% of 100 is – applying a complex algorithm – approximately 20 people. Let’s flag these up on our population schematic in grey.

20 in 100

So 20 people have the wrong diagnosis. One is correctly identified as having PJT Syndrome and 79 are correctly identified as not having PJT Syndrome; so a total of 80 have the right diagnosis.

What does this mean for those 21 people who have been unfortunate enough to test positive for PJT Syndrome (the one person coloured red and the 20 coloured grey)? Well only one of them actually has the malady. So, if I test positive, my chances of actually having PJT Syndrome are not 80% as we originally thought, but instead 1 in 21 or 4.76%. So my risk is still low having tested positive. It is higher than the risk in the general population, which is 1 in 100, or 1%, but not much more so.

The problem arises if having a condition is rare (here 1 in 100) and the accuracy of a test is low (here it is wrong for 20% of people taking it). If you consider that the condition that I was being offered a test for actually has an incidence of around 1 in 20,000 people, then with an 80% accurate test we would get the following:

  1. In a population of 20,000 one 1 person has the condition
  2. In the same population a test with our 80% accuracy means that 20% of people will test positive for it when they are perfectly healthy, this amounts to 4,000 people
  3. So in total, 4,001 people will test positive, 1 correctly, 4,000 erroneously
  4. Which means that a positive test tells me my odds of having the condition being tested for are 1 in 4,001, or 0.025%; still a pretty unlikely event

Low accuracy tests and rare conditions are a very bad combination. As well as causing people unnecessary distress, the real problem is where the diagnosis leads potential suffers to take actions (e.g. undergoing further diagnosis, which could be invasive, or even embarking on a course of treatment) which may themselves have the potential to cause injury to the patient.

I am not of course suggesting that people ignore medical advice, but Doctors are experts in medicine and not statistics. When deciding what course of action to take in a situation similar to one I recently experienced, taking the time to more accurately assess risks and benefits is extremely important. Humans are well known to overestimate some risks (and underestimate others), there are circumstances when crunching the numbers and seeing what they tell you is not only a good idea, it can help to safeguard your health.

For what it’s worth, I opted out of these particular tests.
 


 
Notes

 
[1]
 
A terrible condition which renders sufferers unable to express any thought in under 1,000 words.
 
[2]
 
Not the actual figure quoted, but close to it.

 

 

Curiouser and Curiouser – The Limits of Brexit Voting Analysis

An original illustration from Charles Lutwidge Dodgson's seminal work would have been better, but sadly none such seems to be extant
 
Down the Rabbit-hole

When I posted my Brexit infographic reflecting the age of voters an obvious extension was to add an indication of the number of people in each age bracket who did not vote as well as those who did. This seemed a relatively straightforward task, but actually proved to be rather troublesome (this may be an example of British understatement). Maybe the caution I gave about statistical methods having a large impact on statistical outcomes in An Inconvenient Truth should have led me to expect such issues. In any case, I thought that it would be instructive to talk about the problems I stumbled across and to – once again – emphasise the perils of over-extending statistical models.

Brexit ages infographic
Click to download a larger PDF version in a new window.

Regular readers will recall that my Brexit Infographic (reproduced above) leveraged data from an earlier article, A Tale of two [Brexit] Data Visualisations. As cited in this article, the numbers used were from two sources:

  1. The UK Electoral Commission – I got the overall voting numbers from here.
  2. Lord Ashcroft’s Poling organisation – I got the estimated distribution of votes by age group from here.

In the notes section of A Tale of two [Brexit] Data Visualisations I [prophetically] stated that the breakdown of voting by age group was just an estimate. Based on what I have discovered since, I’m rather glad that I made this caveat explicit.
 
 
The Pool of Tears

In order to work out the number of people in each age bracket who did not vote, an obvious starting point would be the overall electorate, which the UK Electoral Commission stated as being 46,500,001. As we know that 33,551,983 people voted (an actual figure rather than an estimate), then this is where the turnout percentage of 72.2% (actually 72.1548%) came from (33,551,983 / 45,500,001).

A clarifying note, the electorate figures above refer to people who are eligible to vote. Specifically, in order to vote in the UK Referendum, people had to meet the following eligibility criteria (again drawn from the UK Electoral Commission):

To be eligible to vote in the EU Referendum, you must be:

  • A British or Irish citizen living in the UK, or
  • A Commonwealth citizen living in the UK who has leave to remain in the UK or who does not require leave to remain in the UK, or
  • A British citizen living overseas who has been registered to vote in the UK in the last 15 years, or
  • An Irish citizen living overseas who was born in Northern Ireland and who has been registered to vote in Northern Ireland in the last 15 years.

EU citizens are not eligible to vote in the EU Referendum unless they also meet the eligibility criteria above.

So far, so simple. The next thing I needed to know was how the electorate was split by age. This is where we begin to run into problems. One place to start is the actual population of the UK as at the last census (2011). This is as follows:
 

Ages (years) Population % of total
0–4 3,914,000 6.2
5–9 3,517,000 5.6
10–14 3,670,000 5.8
15–19 3,997,000 6.3
20–24 4,297,000 6.8
25–29 4,307,000 6.8
30–34 4,126,000 6.5
35–39 4,194,000 6.6
40–44 4,626,000 7.3
45–49 4,643,000 7.3
50–54 4,095,000 6.5
55–59 3,614,000 5.7
60–64 3,807,000 6.0
65–69 3,017,000 4.8
70–74 2,463,000 3.9
75–79 2,006,000 3.2
80–84 1,496,000 2.4
85–89 918,000 1.5
90+ 476,000 0.8
Total 63,183,000 100.0

 
If I roll up the above figures to create the same age groups as in the Ashcroft analysis (something that requires splitting the 15-19 range, which I have assumed can be done uniformly), I get:
 

Ages (years) Population % of total
0-17 13,499,200 21.4
18-24 5,895,800 9.3
25-34 8,433,000 13.3
35-44 8,820,000 14.0
45-54 8,738,000 13.8
55-64 7,421,000 11.7
65+ 10,376,000 16.4
Total 63,183,000 100.0

 
The UK Government isn’t interested in the views of people under 18[citation needed], so eliminating this row we get:
 

Ages (years) Population % of total
18-24 5,895,800 11.9
25-34 8,433,000 17.0
35-44 8,820,000 17.8
45-54 8,738,000 17.6
55-64 7,421,000 14.9
65+ 10,376,000 20.9
Total 49,683,800 100.0

 
As mentioned, the above figures are from 2011 and the UK population has grown since then. Web-site WorldOMeters offers an extrapolated population of 65,124,383 for the UK in 2016 (this is as at 12th July 2016; if extrapolation and estimates make you queasy, I’d suggest closing this article now!). I’m going to use a rounder figure of 65,125,000 people; there is no point pretending that precision exists where it clearly doesn’t. Making the assumption that such growth is uniform across all age groups (please refer to my previous bracketed comment!), then the above exhibit can also be extrapolated to give us:
 

Ages (years) Population % of total
18-24 6,077,014 11.9
25-34 8,692,198 17.0
35-44 9,091,093 17.8
45-54 9,006,572 17.6
55-64 7,649,093 14.9
65+ 10,694,918 20.9
Total 51,210,887 100.0

 
 
Looking Glass House

So our – somewhat fabricated – figure for the 18+ UK population in 2016 is 51,210,887, let’s just call this 51,200,000. As at the beginning of this article the electorate for the 2016 UK Referendum was 45,500,000 (dropping off the 1 person with apologies to him or her). The difference is explicable based on the eligibility criteria quoted above. I now have a rough age group break down of the 51.2 million population, how best to apply this to the 45.5 million electorate?

I’ll park this question for the moment and instead look to calculate a different figure. Based on the Ashcroft model, what percentage of the UK population (i.e. the 51.2 million) voted in each age group? We can work this one out without many complications as follows:
 

Ages (years)
 
Population
(A)
Voted
(B)
Turnout %
(B/A)
18-24 6,077,014 1,701,067 28.0
25-34 8,692,198 4,319,136 49.7
35-44 9,091,093 5,656,658 62.2
45-54 9,006,572 6,535,678 72.6
55-64 7,649,093 7,251,916 94.8
65+ 10,694,918 8,087,528 75.6
Total 51,210,887 33,551,983 65.5

(B) = Size of each age group in the Ashcroft sample as a percentage multiplied by the total number of people voting (see A Tale of two [Brexit] Data Visualisations).
 
Remember here that actual turnout figures have electorate as the denominator, not population. As the electorate is less than the population, this means that all of the turnout percentages should actually be higher than the ones calculated (e.g. the overall turnout with respect to electorate is 72.2% whereas my calculated turnout with respect to population is 65.5%). So given this, how to explain the 94.8% turnout of 55-64 year olds? To be sure this group does reliably turn out to vote, but did essentially all of them (remembering that the figures in the above table are too low) really vote in the referendum? This seems less than credible.

The turnout for 55-64 year olds in the 2015 General Election has been estimated at 77%, based on an overall turnout of 66.1% (web-site UK Political Info; once more these figures will have been created based on techniques similar to the ones I am using here). If we assume a uniform uplift across age ranges (that “assume” word again!) then one might deduce that an increase in overall turnout from 66.1% to 72.2%, might lead to the turnout in the 55-64 age bracket increasing from 77% to 84%. 84% turnout is still very high, but it is at least feasible; close to 100% turnout in from this age group seems beyond the realms of likelihood.

So what has gone wrong? Well so far the only culprit I can think of is the distribution of voting by age group in the Ashcroft poll. To be clear here, I’m not accusing Lord Ashcroft and his team of sloppy work. Instead I’m calling out that the way that I have extrapolated their figures may not be sustainable. Indeed, if my extrapolation is valid, this would imply that the Ashcroft model over estimated the proportion of 55-64 year olds voting. Thus it must have underestimated the proportion of voters in some other age group. Putting aside the likely fact that I have probably used their figures in an unintended manner, could it be that the much-maligned turnout of younger people has been misrepresented?

To test the validity of this hypothesis, I turned to a later poll by Omnium. To be sure this was based on a sample size of around 2,000 as opposed to Ashcroft’s 12,000, but it does paint a significantly different picture. Their distribution of voter turnout by age group was as follows:
 

Ages (years) Turnout %
18-24 64
25-39 65
40-54 66
55-64 74
65+ 90

 
I have to say that the Omnium age groups are a bit idiosyncratic, so I have taken advantage of the fact that the figures for 25-54 are essentially the same to create a schedule that matches the Ashcroft groups as follows:
 

Ages (years) Turnout %
18-24 64
25-34 65
35-44 65
45-54 65
55-64 74
65+ 90

 
The Omnium model suggests that younger voters may have turned out in greater numbers than might be thought based on the Ashcroft data. In turn this would suggest that a much greater percentage of 18-24 year olds turned out for the Referendum (64%) than for the last General Election (43%); contrast this with an estimated 18-24 turnout figure of 47% based on the just increase in turnout between the General Election and the Referendum. The Omnium estimates do still however recognise that turnout was still greater in the 55+ brackets, which supports the pattern seen in other elections.
 
 
Humpty Dumpty

While it may well be that the Leave / Remain splits based on the Ashcroft figures are reasonable, I’m less convinced that extrapolating these same figures to make claims about actual voting numbers by age group (as I have done) is tenable. Perhaps it would be better to view each age cohort as a mini sample to be treated independently. Based on the analysis above, I doubt that the turnout figures I have extrapolated from the Ashcroft breakdown by age group are robust. However, that is not the same as saying that the Ashcroft data is flawed, or that the Omnium figures are correct. Indeed the Omnium data (at least those elements published on their web-site) don’t include an analysis of whether the people in their sample voted Leave or Remain, so direct comparison is not going to be possible. Performing calculation gymnastics such as using the Omnium turnout for each age group in combination with the Ashcroft voting splits for Leave and Remain for the same age groups actually leads to a rather different Referendum result, so I’m not going to plunge further down this particular rabbit hole.

In summary, my supposedly simple trip to the destitution of an enhanced Brexit Infographic has proved unexpectedly arduous, winding and beset by troubles. These challenges have proved so great that I’ve abandoned the journey and will be instead heading for home.
 
 
Which dreamed it?

Based on my work so far, I have severe doubts about the accuracy of some of the age-based exhibits I have published (versions of which have also appeared on many web-sites, the BBC to offer just one example, scroll down to “How different age groups voted” and note that the percentages cited reconcile to mine). I believe that my logic and calculations are sound, but it seems that I am making too many assumptions about how I can leverage the Ashcroft data. After posting this article, I will accordingly go back and annotate each of my previous posts and link them to these later findings.

I think the broader lesson to be learnt is that estimates are just that, attempts (normally well-intentioned of course) to come up with figures where the actual numbers are not accessible. Sometimes this is a very useful – indeed indispensable – approach, sometimes it is less helpful. In either case estimation should always be approached with caution and the findings ideally sense-checked in the way that I have tried to do above.

Occam’s razor would suggest that when the stats tell you something that seems incredible, then 99 times out of 100 there is an error or inaccurate assumption buried somewhere in the model. This applies when you are creating the model yourself and doubly so where you are relying upon figures calculated by other people. In the latter case not only is there the risk of their figures being inaccurate, there is the incremental risk that you interpret them wrongly, or stretch their broader application to breaking point. I was probably guilty of one or more of the above sins in my earlier articles. I’d like my probable misstep to serve as a warning to other people when they too look to leverage statistics in new ways.

A further point is the most advanced concepts I have applied in my calculations above are addition, subtraction, multiplication and division. If these basic operations – even in the hands of someone like me who is relatively familiar with them – can lead to the issues described above, just imagine what could result from the more complex mathematical techniques (e.g. ambition, distraction, uglification and derision) used by even entry-level data scientists. This perhaps suggests an apt aphorism: Caveat calculator!

Beware the Jabberwock, my son! // The jaws that bite, the claws that catch! // Beware the Jubjub bird, and shun // The frumious Bandersnatch!

 

How Age was a Critical Factor in Brexit

Brexit ages infographic
Click to download a larger PDF version in a new window.

In my last article, I looked at a couple of ways to visualise the outcome of the recent UK Referendum on Europen Union membership. There I was looking at how different visual representations highlight different attributes of data.

I’ve had a lot of positive feedback about my previous Brexit exhibits and I thought that I’d capture the zeitgeist by offering a further visual perspective, perhaps one more youthful than the venerable pie chart; namely an infographic. My attempt to produce one of these appears above and a full-size PDF version is also just a click away.

For caveats on the provenance of the data, please also see the previous article’s notes section.
 

Addendum

I have leveraged age group distributions from the Ascroft Polling organisation to create this exhibits. Other sites – notably the BBC – have done the same and my figures reconcile to the interpretations in other places. However, based on further analysis, I have some reason to think that either there are issues with the Ashcroft data, or that I have leveraged it in ways that the people who compiled it did not intend. Either way, the Ashcroft numbers lead to the conclusion that close to 100% of 55-64 year olds voted in the UK Referendum, which seems very, very unlikely. I have contacted the Ashcroft Polling organisation about this and will post any reply that I receive.

– Peter James Thomas, 14th July 2016

 

A Tale of Two [Brexit] Data Visualisations

Rip, or is that RIP? With apologies to The Economist and acknowledgement to David Bacon.

I’m continuing with the politics and data visualisation theme established in my last post. However, I’ll state up front that this is not a political article. I have assiduously stayed silent [on this blog at least] on the topic of my country’s future direction, both in the lead up to the 23rd June poll and in its aftermath. Instead, I’m going to restrict myself to making a point about data visualisation; both how it can inform and how it can mislead.

Brexit Bar
UK Referendum on EU Membership – Percentage voting by age bracket (see notes)

The exhibit above is my version of one that has appeared in various publications post referendum, both on-line and print. As is referenced, its two primary sources are the UK Electoral Commission and Lord Ashcroft’s polling organisation. The reason why there are two sources rather than one is explained in the notes section below.

With the caveats explained below, the above chart shows the generational divide apparent in the UK Referendum results. Those under 35 years old voted heavily for the UK to remain in the EU; those with ages between 35 and 44 voted to stay in pretty much exactly the proportion that the country as a whole voted to leave; and those over 45 years old voted increasingly heavily to leave as their years advanced.

One thing which is helpful about this exhibit is that it shows in what proportion each cohort voted. This means that the type of inferences I made in the previous paragraph leap off the page. It is pretty clear (visually) that there is a massive difference between how those aged 18-24 and those aged 65+ thought about the question in front of them in the polling booth. However, while the percentage based approach illuminates some things, it masks others. A cursory examination of the chart above might lead one to ask – based on the area covered by red rectangles – how it was that the Leave camp prevailed? To pursue an answer to this question, let’s consider the data with a slightly tweaked version of the same visualisation as below:

Brexit Bar 2
UK Referendum on EU Membership – Numbers voting by age bracket (see notes)

[Aside: The eagle-eyed amongst you may notice a discrepancy between the figures shown on the total bars above and the actual votes cast, which were respectively: Remain: 16,141k and Leave: 17,411k. Again see the notes section for an explanation of this.]

A shift from percentages to actual votes recorded casts some light on the overall picture. It now becomes clear that, while a large majority of 18-24 year olds voted to Remain, not many people in this category actually voted. Indeed while, according to the 2011 UK Census, the 18-24 year category makes up just under 12% of all people over 18 years old (not all of whom would necessarily be either eligible or registered to vote) the Ashcroft figures suggest that well under half of this group cast their ballot, compared to much higher turnouts for older voters (once more see the notes section for caveats).

This observation rather blunts the assertion that the old voted in ways that potentially disadvantaged the young; the young had every opportunity to make their voice heard more clearly, but didn’t take it. Reasons for this youthful disengagement from the political process are of course beyond the scope of this article.

However it is still hard (at least for the author’s eyes) to get the full picture from the second chart. In order to get a more visceral feeling for the dynamics of the vote, I have turned to the much maligned pie chart. I also chose to use the even less loved “exploded” version of this.

Brexit Flag
UK Referendum on EU Membership – Number voting by age bracket (see notes)

Here the weight of both the 65+ and 55+ Leave vote stands out as does the paucity of the overall 18-24 contribution; the only two pie slices too small to accommodate an internal data label. This exhibit immediately shows where the referendum was won and lost in a way that is not as easy to glean from a bar chart.

While I selected an exploded pie chart primarily for reasons of clarity, perhaps the fact that the resulting final exhibit brings to mind a shattered and reassembled Union Flag was also an artistic choice. Unfortunately, it seems that this resemblance has a high likelihood of proving all too prophetic in the coming months and years.
 

Addendum

I have leveraged age group distributions from the Ascroft Polling organisation to create these exhibits. Other sites – notably the BBC – have done the same and my figures reconcile to the interpretations in other places. However, based on further analysis, I have some reason to think that either there are issues with the Ashcroft data, or that I have leveraged it in ways that the people who compiled it did not intend. Either way, the Ashcroft numbers lead to the conclusion that close to 100% of 55-64 year olds voted in the UK Referendum, which seems very, very unlikely. I have contacted the Ashcroft Polling organisation about this and will post any reply that I receive.

– Peter James Thomas, 14th July 2016

 



 
 
Notes

Caveat: I am neither a professional political pollster, nor a statistician. Instead I’m a Pure Mathematician, with a basic understanding of some elements of both these areas. For this reason, the following commentary may not be 100% rigorous; however my hope is that it is nevertheless informative.

In the wake of the UK Referendum on EU membership, a lot of attempts were made to explain the result. Several of these used splits of the vote by demographic attributes to buttress the arguments that they were making. All of the exhibits in this article use age bands, one type of demographic indicator. Analyses posted elsewhere looked at things like the influence of the UK’s social grade classifications (A, B, C1 etc.) on voting patterns, the number of immigrants in a given part of the country, the relative prosperity of different areas and how this has changed over time. Other typical demographic dimensions might include gender, educational achievement or ethnicity.

However, no demographic information was captured as part of the UK referendum process. There is no central system which takes a unique voting ID and allocates attributes to it, allowing demographic dicing and slicing (to be sure a partial and optional version of this is carried out when people leave polling stations after a General Election, but this was not done during the recent referendum).

So, how do so many demographic analyses suddenly appear? To offer some sort of answer here, I’ll take you through how I built the data set behind the exhibits in this article. At the beginning I mentioned that I relied on two data sources, the actual election results published by the UK Electoral Commission and the results of polling carried out by Lord Ashcroft’s organisation. The latter covered interviews with 12,369 people selected to match what was anticipated to be the demographic characteristics of the actual people voting. As with most statistical work, properly selecting a sample with no inherent biases (e.g. one with the same proportion of people who are 65 years or older as in the wider electorate) is generally the key to accuracy of outcome.

Importantly demographic information is known about the sample (which may also be reweighted based on interview feedback) and it is by assuming that what holds true for the sample also holds true for the electorate that my charts are created. So if X% of 18-24 year olds in the sample voted Remain, the assumption is that X% of the total number of 18-24 year olds that voted will have done the same.

12,000 plus is a good sample size for this type of exercise and I have no reason to believe that Lord Ashcroft’s people were anything other than professional in selecting the sample members and adjusting their models accordingly. However this is not the same as having definitive information about everyone who voted. So every exhibit you see relating to the age of referendum voters, or their gender, or social classification is based on estimates. This is a fact that seldom seems to be emphasised by news organisations.

The size of Lord Ashchoft’s sample also explains why the total figures for Leave and Remain on my second exhibit are different to the voting numbers. This is because 5,949 / 12,369 = 48.096% (looking at the sample figures for Remain) whereas 16,141,241 / 33,551,983 = 48.108% (looking at the actual voting figures for Remain). Both figures round to 48.1%, but the small difference in the decimal expansions, when applied to 33 million people, yields a slightly different result.

 

An Inconvenient Truth

Frequentists vs. Bayesians - © xkcd.com
© xkcd.com (adapted from the original to fit the dimensions of this page)

No, not a polemic about climate change, but instead some observations on the influence of statistical methods on statistical findings. It is clearly a truism to state that there are multiple ways to skin a cat, what is perhaps less well-understood is that not all methods of flaying will end up with a cutaneously-challenged feline and some may result in something altogether different.

So an opaque introduction, let me try to shed some light instead. While the points I am going to make here are ones that any statistical practitioner would (or certainly should) know well, they are perhaps less widely appreciated by a general audience. I returned to thinking about this area based on an article by Raphael Silberzahn and Eric Uhlmann in Nature [1], but one which I have to admit first came to my attention via The Economist [2].

Messrs Silberzahn and Uhlmann were propounding a crowd-sourced approach to statistical analysis in science, in particular the exchange of ideas about a given analysis between (potentially rival) groups before conclusions are reached and long before the customary pre- and post-publication reviews. While this idea may well have a lot of merit, I’m instead going to focus on the experiment that the authors performed, some of its results and their implications for more business-focussed analysis teams and individuals.

The interesting idea here was that Silberzahn and Uhlmann provided 29 different teams of researchers the same data set and asked them to investigate the same question. The data set was a sporting one covering the number of times that footballers (association in this case, not American) were dismissed from the field of play by an official. The data set included many attributes from the role of the player, to when the same player / official encountered each other, to demographics of the players themselves. The question was – do players with darker skins get dismissed more often than their fairer teammates?

Leaving aside the socio-political aspects that this problem brings to mind, the question is one that, at least on first glance, looks as if it should be readily susceptible to statistical analysis and indeed the various researchers began to develop their models and tests. A variety of methodologies was employed, “everything from Bayesian clustering to logistic regression and linear modelling” (the authors catalogued the approaches as well as the results) and clearly each team took decisions as to which data attributes were the most significant and how their analyses would be parameterised. Silberzahn and Uhlmann then compared the results.

Below I’ll simply repeat part of their comments (with my highlighting):

Of the 29 teams, 20 found a statistically significant correlation between skin colour and red cards […]. The median result was that dark-skinned players were 1.3 times more likely than light-skinned players to receive red cards. But findings varied enormously, from a slight (and non-significant) tendency for referees to give more red cards to light-skinned players to a strong trend of giving more red cards to dark-skinned players.

This diversity in findings is neatly summarised in the following graph (please click to view the original on Nature’s site):

Nature Graph

© NPG. Used under license 3741390447060 Copyright Clearance Center

To be clear here, the unanimity of findings that one might have expected from analysing what is essentially a pretty robust and conceptually simple data set was essentially absent. What does this mean aside from potentially explaining some of the issues with repeatability that have plagued some parts of science in recent years?

Well the central observation is that precisely the same data set can lead to wildly different insights dependent on how it is analysed. It is not necessarily the case that one method is right and others wrong, indeed in review of the experiment, the various research teams agreed that the approaches taken by others were also valid. Instead it is extremely difficult to disentangle results from the algorithms employed to derive them. In this case methodology had a bigger impact on findings than any message lying hidden in the data.

Here we are talking about leading scientific researchers, whose prowess in statistics is a core competency. Let’s now return to the more quotidian world of the humble data scientist engaged in helping an organisation to take better decisions through statistical modelling. Well the same observations apply. In many cases, insight will be strongly correlated with how the analysis is performed and the choices that the analyst has made. Also, it may not be that there is some objective truth hidden in a dataset, instead only a variety of interpretations of this.

Now this sounds like a call to abandon all statistical models. Nothing could be further from my point of view [3]. However caution is required. In particular those senior business people who place reliance on the output of models, but who maybe do not have a background in statistics, should perhaps ask themselves whether what their organisation’s models tell them is absolute truth, or instead simply more of an indication. They should also ask whether a different analysis methodology might have yielded a different result and thus dictated different business action.

At the risk of coming over all Marvel, the great power of statistical modelling comes with great responsibility.

In 27 years in general IT and 15 in the data/information space (to say nothing of my earlier Mathematical background) I have not yet come across a silver bullet. My strong suspicion is that they don’t exist. However, I’d need to carry out some further analysis to reach a definitive conclusion; now what methodology to employ…?
 


 
Notes

 
[1]
 
Crowdsourced research: Many hands make tight work. Raphael Silberzahn &a Eric L. Uhlmann. Nature. 07 October 2015.
07 October 2015
 
[2]
 
On the other hands – Honest disagreement about methods may explain irreproducible results.The Economist 10th Oct 2015.
 
[3]
 
See the final part of my trilogy on using historical data to justify BI investments for a better representation of my actual views.

Data Visualisation – A Scientific Treatment

Introduction

Diagram of the Causes of Mortality in the Army of the East (click to view a larger version in a new tab)

The above diagram was compiled by Florence Nightingale, who was – according to The Font – “a celebrated English social reformer and statistician, and the founder of modern nursing”. It is gratifying to see her less high-profile role as a number-cruncher acknowledged up-front and central; particularly as she died in 1910, eight years before women in the UK were first allowed to vote and eighteen before universal suffrage. This diagram is one of two which are generally cited in any article on Data Visualisation. The other is Charles Minard’s exhibit detailing the advance on, and retreat from, Moscow of Napoleon Bonaparte’s Grande Armée in 1812 (Data Visualisation had a military genesis in common with – amongst many other things – the internet). I’ll leave the reader to look at this second famous diagram if they want to; it’s just a click away.

While there are more elements of numeric information in Minard’s work (what we would now call measures), there is a differentiating point to be made about Nightingale’s diagram. This is that it was specifically produced to aid members of the British parliament in their understanding of conditions during the Crimean War (1853-56); particularly given that such non-specialists had struggled to understand traditional (and technical) statistical reports. Again, rather remarkably, we have here a scenario where the great and the good were listening to the opinions of someone who was barred from voting on the basis of lacking a Y chromosome. Perhaps more pertinently to this blog, this scenario relates to one of the objectives of modern-day Data Visualisation in business; namely explaining complex issues, which don’t leap off of a page of figures, to busy decision makers, some of whom may not be experts in the specific subject area (another is of course allowing the expert to discern less than obvious patterns in large or complex sets of data). Fortunately most business decision makers don’t have to grapple with the progression in number of “deaths from Preventible or Mitigable Zymotic diseases” versus ”deaths from wounds” over time, but the point remains.
 
 
Data Visualisation in one branch of Science

von Laue, Bragg Senior & Junior, Crowfoot Hodgkin, Kendrew, Perutz, Crick, Franklin, Watson & Wilkins

Coming much more up to date, I wanted to consider a modern example of Data Visualisation. As with Nightingale’s work, this is not business-focused, but contains some elements which should be pertinent to the professional considering the creation of diagrams in a business context. The specific area I will now consider is Structural Biology. For the incognoscenti (no advert for IBM intended!), this area of science is focussed on determining the three-dimensional shape of biologically relevant macro-molecules, most frequently proteins or protein complexes. The history of Structural Biology is intertwined with the development of X-ray crystallography by Max von Laue and father and son team William Henry and William Lawrence Bragg; its subsequent application to organic molecules by a host of pioneers including Dorothy Crowfoot Hodgkin, John Kendrew and Max Perutz; and – of greatest resonance to the general population – Francis Crick, Rosalind Franklin, James Watson and Maurice Wilkins’s joint determination of the structure of DNA in 1953.

photo-51

X-ray diffraction image of the double helix structure of the DNA molecule, taken 1952 by Raymond Gosling, commonly referred to as “Photo 51”, during work by Rosalind Franklin on the structure of DNA

While the masses of data gathered in modern X-ray crystallography needs computer software to extrapolate them to physical structures, things were more accessible in 1953. Indeed, it could be argued that Gosling and Franklin’s famous image, its characteristic “X” suggestive of two helices and thus driving Crick and Watson’s model building, is another notable example of Data Visualisation; at least in the sense of a picture (rather than numbers) suggesting some underlying truth. In this case, the production of Photo 51 led directly to the creation of the even more iconic image below (which was drawn by Francis Crick’s wife Odile and appeared in his and Watson’s seminal Nature paper[1]):

Odile and Francis Crick - structure of DNA

© Nature (1953)
Posted on this site under the non-commercial clause of the right-holder’s licence

It is probably fair to say that the visualisation of data which is displayed above has had something of an impact on humankind in the fifty years since it was first drawn.
 
 
Modern Structural Biology

The X-ray Free Electron Laser at Stanford

Today, X-ray crystallography is one of many tools available to the structural biologist with other approaches including Nuclear Magnetic Resonance Spectroscopy, Electron Microscopy and a range of biophysical techniques which I will not detain the reader by listing. The cutting edge is probably represented by the X-ray Free Electron Laser, a device originally created by repurposing the linear accelerators of the previous generation’s particle physicists. In general Structural Biology has historically sat at an intersection of Physics and Biology.

However, before trips to synchrotrons can be planned, the Structural Biologist often faces the prospect of stabilising their protein of interest, ensuring that they can generate sufficient quantities of it, successfully isolating the protein and finally generating crystals of appropriate quality. This process often consumes years, in some cases decades. As with most forms of human endeavour, there are few short-cuts and the outcome is at least loosely correlated to the amount of time and effort applied (though sadly with no guarantee that hard work will always be rewarded).
 
 
From the general to the specific

The Journal of Molecular Biology (October 2014)

At this point I should declare a personal interest, the example of Data Visualisation which I am going to consider is taken from a paper recently accepted by the Journal of Molecular Biology (JMB) and of which my wife is the first author[2]. Before looking at this exhibit, it’s worth a brief detour to provide some context.

In recent decades, the exponential growth in the breadth and depth of scientific knowledge (plus of course the velocity with which this can be disseminated), coupled with the increase in the range and complexity of techniques and equipment employed, has led to the emergence of specialists. In turn this means that, in a manner analogous to the early production lines, science has become a very collaborative activity; expert in stage one hands over the fruits of their labour to expert in stage two and so on. For this reason the typical scientific paper (and certainly those in Structural Biology) will have several authors, often spread across multiple laboratory groups and frequently in different countries. By way of example the previous paper my wife worked on had 16 authors (including a Nobel Laureate[3]). In this context, the fact the paper I will now reference was authored by just my wife and her group leader is noteworthy.

The reader may at this point be relieved to learn that I am not going to endeavour to explain the subject matter of my wife’s paper, nor the general area of biology to which it pertains (the interested are recommended to Google “membrane proteins” or “G Protein Coupled Receptors” as a starting point). Instead let’s take a look at one of the exhibits.

Click to view a larger version in a new tab

© The Journal of Molecular Biology (2014)
Posted on this site under a Creative Commons licence

The above diagram (in common with Nightingale’s much earlier one) attempts to show a connection between sets of data, rather than just the data itself. I’ll elide the scientific specifics here and focus on more general issues.

First the grey upper section with the darker blots on it – which is labelled (a) – is an image of a biological assay called a Western Blot (for the interested, details can be viewed here); each vertical column (labelled at the top of the diagram) represents a sub-experiment on protein drawn from a specific sample of cells. The vertical position of a blot indicates the size of the molecules found within it (in kilodaltons); the intensity of a given blot indicates how much of the substance is present. Aside from the headings and labels, the upper part of the figure is a photographic image and so essentially analogue data[4]. So, in summary, this upper section represents the findings from one set of experiments.

At the bottom – and labelled (b) – appears an artefact familiar to anyone in business, a bar-graph. This presents results from a parallel experiment on samples of protein from the same cells (for the interested, this set of data relates to degree to which proteins in the samples bind to a specific radiolabelled ligand). The second set of data is taken from what I might refer to as a “counting machine” and is thus essentially digital. To be 100% clear, the bar chart is not a representation of the data in the upper part of the diagram, it pertains to results from a second experiment on the same samples. As indicated by the labelling, for a given sample, the column in the bar chart (b) is aligned with the column in the Western Blot above (a), connecting the two different sets of results.

Taken together the upper and lower sections[5] establish a relationship between the two sets of data. Again I’ll skip on the specifics, but the general point is that while the Western Blot (a) and the binding assay (b) tell us the same story, the Western Blot is a much more straightforward and speedy procedure. The relationship that the paper establishes means that just the Western Blot can be used to perform a simple new assay which will save significant time and effort for people engaged in the determination of the structures of membrane proteins; a valuable new insight. Clearly the relationships that have been inferred could equally have been presented in a tabular form instead and be just as relevant. It is however testament to the more atavistic side of humans that – in common with many relationships between data – a picture says it more surely and (to mix a metaphor) more viscerally. This is the essence of Data Visualisation.
 
 
What learnings can Scientific Data Visualisation provide to Business?

Scientific presentation (c/o Nature, but looks a lot like PhD Comics IMO)

Using the JMB exhibit above, I wanted to now make some more general observations and consider a few questions which arise out of comparing scientific and business approaches to Data Visualisation. I think that many of these points are pertinent to analysis in general.
 
 
Normalisation

Broadly, normalisation[6] consists of defining results in relation to some established yardstick (or set of yardsticks); displaying relative, as opposed to absolute, numbers. In the JMB exhibit above, the amount of protein solubilised in various detergents is shown with reference to the un-solubilised amount found in native membranes; these reference figures appear as 100% columns to the right and left extremes of the diagram.

The most common usage of normalisation in business is growth percentages. Here the fact that London business has grown by 5% can be compared to Copenhagen having grown by 10% despite total London business being 20-times the volume of Copenhagen’s. A related business example, depending on implementation details, could be comparing foreign currency amounts at a fixed exchange rate to remove the impact of currency fluctuation.

Normalised figures are very typical in science, but, aside from the growth example mentioned above, considerably less prevalent in business. In both avenues of human endeavour, the approach should be used with caution; something that increases 200% from a very small starting point may not be relevant, be that the result of an experiment or weekly sales figures. Bearing this in mind, normalisation is often essential when looking to present data of different orders on the same graph[7]; the alternative often being that smaller data is swamped by larger, not always what is desirable.
 
 
Controls

I’ll use an anecdote to illustrate this area from a business perspective. Imagine an organisation which (as you would expect) tracks the volume of sales of a product or service it provides via a number of outlets. Imagine further that it launches some sort of promotion, perhaps valid only for a week, and notices an uptick in these sales. It is extremely tempting to state that the promotion has resulted in increased sales[8].

However this cannot always be stated with certainty. Sales may have increased for some totally unrelated reason such as (depending on what is being sold) good or bad weather, a competitor increasing prices or closing one or more of their comparable outlets and so on. Equally perniciously, the promotion maybe have simply moved sales in time – people may have been going to buy the organisation’s product or service in the weeks following a promotion, but have brought the expenditure forward to take advantage of it. If this is indeed the case, an uptick in sales may well be due to the impact of a promotion, but will be offset by a subsequent decrease.

In science, it is this type of problem that the concept of control tests is designed to combat. As well as testing a result in the presence of substance or condition X, a well-designed scientific experiment will also be carried out in the absence of substance or condition X, the latter being the control. In the JMB exhibit above, the controls appear in the columns with white labels.

There are ways to make the business “experiment” I refer to above more scientific of course. In retail business, the current focus on loyalty cards can help, assuming that these can be associated with the relevant transactions. If the business is on-line then historical records of purchasing behaviour can be similarly referenced. In the above example, the organisation could decide to offer the promotion at only a subset of the its outlets, allowing a comparison to those where no promotion applied. This approach may improve rigour somewhat, but of course it does not cater for purchases transferred from a non-promotion outlet to a promotion one (unless a whole raft of assumptions are made). There are entire industries devoted to helping businesses deal with these rather messy scenarios, but it is probably fair to say that it is normally easier to devise and carry out control tests in science.

The general take away here is that a graph which shows some change in a business output (say sales or profit) correlated to some change in a business input (e.g. a promotion, a new product launch, or a price cut) would carry a lot more weight if it also provided some measure of what would have happened without the change in input (not that this is always easy to measure).
 
 
Rigour and Scrutiny

I mention in the footnotes that the JMB paper in question includes versions of the exhibit presented above for four other membrane proteins, this being in order to firmly establish a connection. Looking at just the figure I have included here, each element of the data presented in the lower bar-graph area is based on duplicated or triplicated tests, with average results (and error bars – see the next section) being shown. When you consider that upwards of three months’ preparatory work could have gone into any of these elements and that a mistake at any stage during this time would have rendered the work useless, some impression of the level of rigour involved emerges. The result of this assiduous work is that the authors can be confident that the exhibits they have developed are accurate and will stand up to external scrutiny. Of course such external scrutiny is a key part of the scientific process and the manuscript of the paper was reviewed extensively by independent experts before being accepted for publication.

In the business world, such external scrutiny tends to apply most frequently to publicly published figures (such as audited Financial Accounts); of course external financial analysts also will look to dig into figures. There may be some internal scrutiny around both the additional numbers used to run the business and the graphical representations of these (and indeed some companies take this area very seriously), but not every internal KPI is vetted the way that the report and accounts are. Particularly in the area of Data Visualisation, there is a tension here. Graphical exhibits can have a lot of impact if they relate to the current situation or present trends; contrawise if they are substantially out-of-date, people may question their relevance. There is sometimes the expectation that a dashboard is just like its aeronautical counterpart, showing real-time information about what is going on now[9]. However a lot of the value of Data Visualisation is not about the here and now so much as trends and explanations of the factors behind the here and now. A well-thought out graph can tell a very powerful story, more powerful for most people than a table of figures. However a striking graph based on poor quality data, data which has been combined in the wrong way, or even – as sometimes happens – the wrong datasets entirely, can tell a very misleading story and lead to the wrong decisions being taken.

I am not for a moment suggesting here that every exhibit produced using Data Visualisation tools must be subject to months of scrutiny. As referenced above, in the hands of an expert such tools have the value of sometimes quickly uncovering hidden themes or factors. However, I would argue that – as in science – if the analyst involved finds something truly striking, an association which he or she feels will really resonate with senior business people, then double- or even triple-checking the data would be advisable. Asking a colleague to run their eye over the findings and to then probe for any obvious mistakes or weaknesses sounds like an appropriate next step. Internal Data Visualisations are never going to be subject to peer-review, however their value in taking sound business decisions will be increased substantially if their production reflects at least some of the rigour and scrutiny which are staples of the scientific method.
 
 
Dealing with Uncertainty

In the previous section I referred to the error bars appearing on the JMB figure above. Error bars are acknowledgements that what is being represented is variable and they indicate the extent of such variability. When dealing with a physical system (be that mechanical or – as in the case above – biological), behaviour is subject to many factors, not all of which can be eliminated or adjusted for and not all of which are predictable. This means that repeating an experiment under ostensibly identical conditions can lead to different results[10]. If the experiment is well-designed and if the experimenter is diligent, then such variability is minimised, but never eliminated. Error bars are a recognition of this fundamental aspect of the universe as we understand it.

While de rigueur in science, error bars seldom make an appearance in business, even – in my experience – in estimates of business measures which emerge from statistical analyses[11]. Even outside the realm of statistically generated figures, more business measures are subject to uncertainty than might initially be thought. An example here might be a comparison (perhaps as part of the externally scrutinised report and accounts) of the current quarter’s sales to the previous one (or the same one last year). In companies where sales may be tied to – for example – the number of outlets, care is paid to making these figures like-for-like. This might include only showing numbers for outlets which were in operation in the prior period and remain in operation now (i.e. excluding sales from both closed outlets or newly opened ones). However, outside the area of high-volume low-value sales where the Law of Large Numbers[12] rules, other factors could substantially skew a given quarter’s results for many organisations. Something as simple as a key customer delaying a purchase (so that it fell in Q3 this year instead of Q2 last) could have a large impact on quarterly comparisons. Again companies will sometimes look to include adjustments to cater for such timing or related issues, but this cannot be a precise process.

The main point I am making here is that many aspects of the information produced in companies is uncertain. The cash transactions in a quarter are of course the cash transactions in a quarter, but the above scenario suggests that they may not always 100% reflect actual business conditions (and you cannot adjust for everything). Equally where you get in to figures that would be part of most companies’ financial results, outstanding receivables and allowance for bad debts, the spectre of uncertainty arises again without a statistical model in sight. In many industries, regulators are pushing for companies to include more forward-looking estimates of future assets and liabilities in their Financials. While this may be a sensible reaction to recent economic crises, the approach inevitably leads to more figures being produced from models. Even when these models are subject to external review, as is the case with most regulatory-focussed ones, they are still models and there will be uncertainty around the numbers that they generate. While companies will often provide a range of estimates for things like guidance on future earnings per share, providing a range of estimates for historical financial exhibits is not really a mainstream activity.

Which perhaps gets me back to the subject of error bars on graphs. In general I think that their presence in Data Visualisations can only add value, not subtract it. In my article entitled Limitations of Business Intelligence I include the following passage which contains an exhibit showing how the Bank of England approaches communicating the uncertainty inevitably associated with its inflation estimates:

Business Intelligence is not a crystal ball, Predictive Analytics is not a crystal ball either. They are extremely useful tools […] but they are not universal panaceas.

The Old Lady of Threadneedle Street is clearly not a witch
An inflation prediction from The Bank of England
Illustrating the fairly obvious fact that uncertainty increases in proportion to time from now.
[…] Statistical models will never give you precise answers to what will happen in the future – a range of outcomes, together with probabilities associated with each is the best you can hope for (see above). Predictive Analytics will not make you prescient, instead it can provide you with useful guidance, so long as you remember it is a prediction, not fact.

While I can’t see them figuring in formal financial statements any time soon, perhaps there is a case for more business Data Visualisations to include error bars.
 
 
In Summary

So, as is often the case, I have embarked on a journey. I started with an early example of Data Visualisation, diverted in to a particular branch of science with which I have some familiarity and hopefully returned, again as is often the case, to make some points which I think are pertinent to both the Business Intelligence practitioner and the consumers (and indeed commissioners) of Data Visualisations. Back in “All that glisters is not gold” – some thoughts on dashboards I made some more general comments about the best Data Visualisations having strong informational foundations underpinning them. While this observation remains true, I do see a lot of value in numerically able and intellectually curious people using Data Visualisation tools to quickly make connections which had not been made before and to tease out patterns from large data sets. In addition there can be great value in using Data Visualisation to present more quotidian information in a more easily digestible manner. However I also think that some of the learnings from science which I have presented in this article suggest that – as with all powerful tools – appropriate discretion on the part of the people generating Data Visualisation exhibits and on the part of the people consuming such content would be prudent. In particular the business equivalents of establishing controls, applying suitable rigour to data generation / combination and including information about uncertainty on exhibits where appropriate are all things which can help make Data Visualisation more honest and thus – at least in my opinion – more valuable.
 


 
Notes

 
[1]
 
Watson, J.D., Crick, F.H.C. (1953). Molecular structure of nucleic acids; a structure for deoxyribose nucleic acid. Nature.
 
[2]
 
Thomas, J.A., Tate, C.G. (2014). Quality Control in Eukaryotic Membrane Protein Overproduction. J. Mol. Biol. [Epub ahead of print].
 
[3]
 
The list of scientists involved in the development of X-ray Crystallography and Structural Biology which was presented earlier in the text encompasses a further nine such laureates (four of whom worked at my wife’s current research institute), though sadly this number does not include Rosalind Franklin. Over 20 Nobel Prizes have been awarded to people working in the field of Structural Biology, you can view an interactive time line of these here.
 
[4]
 
The intensity, size and position of blots are often digitised by specialist software, but this is an aside for our purposes.
 
[5]
 
Plus four other analogous exhibits which appear in the paper and relate to different proteins.
 
[6]
 
Normalisation has a precise mathematical meaning, actually (somewhat ironically for that most precise of activities) more than one. Here I am using the term more loosely.
 
[7]
 
That’s assuming you don’t want to get into log scales, something I have only come across once in over 25 years in business.
 
[8]
 
The uptick could be as compared to the week before, or to some other week (e.g. the same one last year or last month maybe) or versus an annual weekly average. The change is what is important here, not what the change is with respect to.
 
[9]
 
Of course some element of real-time information is indeed both feasible and desirable; for more analytic work (which encompasses many aspects of Data Visualisation) what is normally more important is sufficient historical data of good enough quality.
 
[10]
 
Anyone interested in some of the reasons for this is directed to my earlier article Patterns patterns everywhere.
 
[11]
 
See my series of three articles on Using historical data to justify BI investments for just one example of these.
 
[12]
 
But then 1=2 for very large values of 1