More on how Big Data didn't win it

Buzzfeed published an article back in February about the 'Big Data' thing and Donald Trump's campaign:

Several people who worked directly with Cambridge Analytica told BuzzFeed News that despite its sales pitch and public statements, it never provided any proof that the technique was effective or that the company had the ability to execute it on a large scale. “Anytime we ever wanted to test anything as far as psychographic was concerned, they would get very hesitant,” said one former campaign staffer. “At no point did they provide us any documentation that it would work.”

Now I suppose that, if you're a conspiracy theorist, this can be dismissed as the campaign people throwing chaff out the back of the plane but it reinforces for me the weaknesses of the Cambridge Analytica claims. The article sets out some of these claims - or at least the ones that the company has broadcast proudly - including the use of 'psychographics'.

We know (because Cambridge Analytical tell us so) that the company has a 'psychographic profile' derived either from, questionably-sourced, Facebook data or else the company's own surveys. For this to be usable it has to be translated into a system that can be applied to the whole population (or at least that population in very small units). It seems that the Cambridge Analytica CEO is saying they have such a system because he claims the company had “profiled the personality of every adult in the United States of America — 220 million people.”

The real question here isn't whether or not such a profile exists (and if it is based solely on Facebook it doesn't since over 40% of Americans aren't using Facebook and a large part of those who are are irregular or infrequent users even before privacy choices are considered) but whether it is in any way either meaningful or effective as a marketing tool. I've a feeling that it is little better than standard geodemographics built on Richard Webber's 'birds of a feather flock together' principle. The observations of people involved in the Trump campaign seem to bear this out - Cambridge Analytica's work perhaps did provide useful targeting insights but didn't then provide any useable means of directly reaching these new targets.

In marketing pitches, two GOP operatives recalled, Nix has claimed his company has access to proprietary information that includes Facebook data. One of the operatives said the data was too old to be helpful and couldn’t be updated. Others said they’d received a similar pitch, but Nix was too vague about the details for them to evaluate what the data really was. None of the campaign staffers BuzzFeed News spoke with said Cambridge Analytica’s proprietary data had played a key role in any decision-making.

There may indeed be some strategic value in extensive data analysis but it is not the salvation. Indeed, there's some suggestion that Hillary Clinton's campaign was even more driven by data analytics than Trump's campaign:

What Ada did, based on all that data, aides said, was run 400,000 simulations a day of what the race against Trump might look like. A report that was spit out would give campaign manager Robby Mook and others a detailed picture of which battleground states were most likely to tip the race in one direction or another — and guide decisions about where to spend time and deploy resources.

The use of analytics by campaigns was hardly unprecedented. But Clinton aides were convinced their work, which was far more sophisticated than anything employed by President Obama or GOP nominee Mitt Romney in 2012, gave them a big strategic advantage over Trump.

And we know what happened here - Clinton targeted either places she was winning easily or places where chaotic information led the campaign to believe it was doing better than it was. There seems to have been a deal more art in the Trump campaign than the conspiracy theorists want, with those involved using data more cautiously than was the case for the Democrats. Just as with the EU referendum, we see people wanting to deny the failings of their own campaign by suggesting the matter was determined by devious, sinister, billionaire-funded legerdemain. This problem continues today and, if anything, is expanding as Russian bots, Macedonian fake news and manipulative 'right wing' media are added to the things blamed for the 'wrong' result.

The same message needs to go to Democrats as I sent to Remain supporters in the UK:

At some point the rump of disappointed remain voters will stop trying to find some sinister external force - Russians, American data companies, Facebook - that explains why we voted to leave and recognise that, in truth, we voted to leave because the EU is a distant, anonymous, unapproachable, corrupt and interfering undemocratic institution. That's it - all of it. And if you ask people a slightly different question, they'll tell you that London is also a distant, anonymous, unapproachable, corrupt and interfering undemocratic place too. One run by and for people with more connection to New York or Paris than Barnsley or Stoke. Perhaps those still angst-ridden by us leaving can begin to learn this?

....

An Evil Marketer comments on data analytics in political campaigning

Back in the 1970s a man called Richard Webber was working at the UK government's centre for Environmental Studies. This, more or less, was what he was doing:

I created Acorn, the first neighbourhood classification system, while working at the government's Centre for Environmental Studies in the 1970s.

People ask how it came to be a commercial application. I had organised a seminar for local authorities to show how neighbourhood data could identify areas of deprivation. Quite by chance, it attracted Ken Baker, a sampling specialist from BMRB. He was a lateral thinker and he came up with the idea that the Acorn tool would be useful for market research sampling.

I meanwhile had realised that Acorn could help predict the households most likely to respond to direct mail and door drops.

Without wanting to come over all geeky, Webber had used census data combined with the electoral register to create a classification of small neighbourhoods (based on census enumeration districts, the base geography for the national census). The principle of Webber's classification is essentially that 'birds of a feather flock together' - people in Manchester who have the same census characteristics as people in Bristol will tend to have other similar behavioural characteristics including, of course, purchasing preferences.

I know you're asking what all this has to do with Brexit (or indeed voting in general). The thing is that, as well as Webber's classification of residential neighbourhoods capturing similarities in terms of how people might respond to different marketing offers, the system also applies to the matter of how we vote. If the 'birds of a feather' principle is correct then similar areas will have similar voting behaviour.

By the end of the 1980s, we were using geodemographics (as Webber's system became known) to combine with proprietary data on customers to refine the targeting of direct marketing campaign, to improve the selection of retail sites and to manage advertising better. Without wanting to overcomplicate, customer address data was given an ACORN code and then profiled using initially a simple index (where 100 = National Average). The indices were reported at the level of postal sector (i.e. BD11 2 or ME2 7) as these were large enough to give the index validity but small enough to facilitate fine grain targeting.

For a targeted doordrop we might then take the top 300 postal sectors and, through the Royal Mail, purchase a delivery of unaddressed mail. Typically, results showed an uplift in response of 2x or 3x depending on the client. Results were less good for targeted direct mail using the electoral register - mostly because a key behavioural characteristic, responsiveness, was not captured in census data. We played other games using expert systems and the early days of data-mining too - it was just a lot slower back then!


The generation of geodemographic systems that followed the original census-based ACORN system (e.g. SuperProfiles, MOSAIC) began to add in other large data sources such as credit data and 'psychographic' survey data - some may recall the retailed questionnaires incentivised with free prize draws that collected this information. Targeting extended beyond shared residential characteristics to details about financial services, indebtedness and preferences around holidays, cars, fmcg products and lifestyle choices (smoking, drinking, gambling, etc.).

Unsurprisingly, political parties began to make use of these systems to improve the targeting of election campaigning especially in areas where they lacked good quality 'voting intention' (VI) data. The same essential methodology was used as that I was using for mail order and financial services companies - a database of VI details was profiles against MOSAIC to provide improved campaign targeting in places with no canvass. This might be used on a national basis to decide which local council by-elections to target or at the constituency level to improve the effectiveness of limited resources thereby allowing the broadening of target seat campaigns.

Which I guess brings us to this conclusion from an article about a marketing analytics business who may or may not have been active during the recent EU Referendum:

Is it the case that our elections will increasingly be decided by the whims of billionaires, operating in the shadows, behind the scenes, using their fortunes to decide our fate?

To appreciate why this is unlikely, we need to go back to how marketing analytics work, which is at the aggregate level. I appreciate, as a marketing professional, how we all want to believe advertising has become like the opening scenes in Minority Report but the reality is that aggregate data really doesn't provide the means to manipulate what we think. Rather, geodemographics, psychographics and marketing analytics enable us marketers to better target those people who are already predisposed to buy our product (or vote for our cause).

The big change from the stuff we were doing with mag tapes and mainframe computers in 1989 is the availability of information from social media (most usually Facebook). Again this is aggregate data - Facebook doesn't sell your details to marketing analytics businesses - but the sophistication of the data makes the targeting of messages all the more precise as does the ability to analyse public profiles without Facebook's permission. But even with this level of analysis, we still haven't 'manipulated' your opinion merely targeted our message more precisely to people more likely to respond positively to that message.

We are right to express concerns about whether the use of analytics has crossed over into misuse of personal data and it appears the UK's Information Commissioner is doing just that. But the likely truth (given it is not in Facebook's interest to share personal data and this is illegal in the EU) is that analytics companies are simply doing just what we were doing as direct marketers 30 years ago - using information to target our messages a little better. Back then it was all seen as a bit sinister ('where did you get my name from...') and nothing has changed except that there's more information, faster computers and social media. You can always find a computing academic (note, not a marketer) do do the full on evil empire stuff:

A rapid convergence in the data mining, algorithmic and granular analytics capabilities of companies like Cambridge Analytica and Facebook is creating powerful, unregulated and opaque ‘intelligence platforms’. In turn, these can have enormous influence to affect what we learn, how we feel, and how we vote. The algorithms they may produce are frequently hidden from scrutiny and we see only the results of any insights they might choose to publish.”

The truth is a deal more prosaic. Marketers ask customers about their lives, social media use and so forth then profile this against aggregate data from various sources to produce targeting information - where geographically or behaviourally we can go looking for folk like those customers we've surveyed. I so want us marketers and ad men to be master manipulators, able to switch your mind at the push of an analytical button or the twitch of an algorhithm. But we're not like that at all, we're not even that good at using big data.

.....

The Metis project won't solve the Conservatives' campaigning problem...

****

There's a belief that Barak Obama won two presidential elections because his team were super slick with the on-line campaigning. All those voter demographics and behavioural metrics were the thing that meant Republicans stood no chance.

So it's no surprise that this things are now all the rage here in the UK - here's Sebastian Payne drooling over one such system in the Spectator:

... an alternative is revealed with the Metis project. Headed up by four of Westminster’s sharpest minds, Metis is destined to become the largest and most sophisticated voter database ever built in the UK. The power of a 20 million strong list of voters has the potential to revolutionise campaigning.

And it will do this by enabling:

...political parties to run highly targeted campaigns, focusing on individual voters whose support is vital to win key seats. More importantly, it will spare householders the sort of unwelcome attention that was lavished on them by over-enthusiastic (or desperate) campaigners in Eastleigh’

This is great - it reminds me of the Asimov short story, "Franchise", where

...the computer Multivac selects a single person to answer a number of questions. Multivac will then use the answers and other data to determine what the results of an election would be, avoiding the need for an actual election to be held.

Such speculation aside, this sophisticated and targeted approach is only half the story of Obama's success - the other have is the activist, the boots on the ground:

So it was that Bird and his colleagues drew up plans to ­expand the electorate into one that could reelect Obama. In Ohio, for example, a “barber shop and beauty salon” strategy was designed to get likely Obama supporters, particularly African-Americans, to register to vote when they went for a haircut. “Faith captains” were assigned to churches to encourage parishioners to turn out for Obama. “Condo captains” were told to know every potential Obama voter in their building. The goal was like nothing seen in presidential politics: Each Obama worker would be ­responsible for about 50 voters in key precincts over the course of the campaign. By Election Day, that worker would know much about the lives of those 50 voters, including whether they had made it to the polls. Romney’s team talked about a ratio of thousands of voters per worker. It would prove to be a crucial difference.

Here lies the other half of the secret - the database that Obama's team used wasn't some clever piece of geodemographics spliced with a lifestyle database and based on questionnaire data. What they were using was real information about real people - and the contact was direct, personal and on the doorstep (or the barber's chair).

If UK political parties think that the solution is to echo Howard Dean's campaign, they are wrong. That campaign failed because it thought that political engagement on-line was everything - it wasn't and it isn't. If we run campaigns on the basis of manipulating large data sets the result will be a worse politics. And for those campaigners the approach probably won't work. Indeed, as Vince-Wayne Mitchell demonstrated years ago, you can make a large data set say almost anything you want it to say:

Suggests that a prima facie case exists for the suitability of astrology as a segmentation variable with the potential to combine the measurement advantages of demographics with the psychological insights of psychographics and to create segments which are measurable, substantial, exhaustive, stable over time, and relatively accessible. Tests the premise empirically using results from a Government data set, the British General Household Survey. The analyses show that astrology does have a significant, and sometimes predictable, effect on behavior in the leisure, tobacco, and drinks markets.

If political parties want to win they need to put boots on the ground, to collect data on the doorstep - for sure the sort of information in Metis will be useful, just as geodemographics have always been useful. To profile, to assist in targeting and to select geographically. These are relevant to politics but, just as is the case with regular marketing, a list of previous buyers - or previous voters - is much more responsive.

The task is to build that list - that is what Obama did. He did use a clever marketing database but applied on-line techniques to the age old method - speak to the voter, look him in the eye and ass; "will you vote for me?"

....

Some tips on geodemographics for Shelter...

****

Back in the days before the PC became ubiquitous, in a time when direct marketing was direct marketing, a new science arrived. It was called geodemographics and we (by which I mean planners in direct marketing agencies) were smitten, deeply smitten. This was the Eldorado of targeted marketing, the Shangri la of response advertising, the holy grail of direct mail. And we played with it – we created mailing lists by ‘profiling’ the electoral register, we segmented and targeted the big agglomerations of address data on the files of big finance and retail businesses and we linked the base geodemographics, the idea that ‘birds of a feather flock together’, to other information collected about customers.

And we learned a couple of important lessons really quickly (mostly through the principle of ‘test and learn’ that direct marketers, uniquely among the marketing breeds, apply to all their work):

1. Geodemographics, for all its wonderfulness is a pretty blunt targeting tool – yes it improves on the random selection from the population but only very slightly. Only where the profile indicated that a given place, typically a postcode area, was at least five times more likely to contain people with a given behaviour did we consider it worthy of selection. And that was for a door drop not for expensive direct mail – for that we needed ten times at least.
2.  For some products and services there simply wasn’t much evidence – beyond confirming existing income-based differences – of a significant variation from par to mean that the ACORN, MOSAIC or Superprofiles analysis. Too often I tore open with rising excitement the envelope containing the profile for a client only to see a flat profile of mostly academic value. To make things work for the client we turned back to old fashioned techniques – using known responders, previous customers and incentivised two-stage campaigns to get new buyers.

So I smile when I see the latest super clever system – using ACORN, MOSAIC or some other geodemographic profiling system as a tool for marketing, communications or even ‘activism’. And the smile is a little wan since these approaches are misleading (to say the least) and of little real value. Here’s an example from housing charity, Shelter:

The Shelter Housing Insights for Communities resource is a must-have for anyone involved in community consultation on housing development. Built using ACORN and extensive bespoke national surveys on housing attitudes, this resource is a unique insight into the housing views and aspirations of local communities; providing advice on cost-effective and targeted consultation.

Using this system you can pop in a postcode and it will tell you whether the local folk are (or maybe aren’t) NIMBYs or BANANAs. Or at least you can at the “ward level” – a level that, in Bradford, is about as much use for getting any understanding as the proverbial chocolate fireguard. I can get a postcode assessment – mine is 3I33 (comfortably off settled suburbia, middle income couples) – but have to track back into Shelter's national assessment to make any sense at all of the information.

And then we learn the bleedin’ obvious – people who live in poor housing especially in cities are supportive of new housing development whereas folk living in suburbia (and especially the wealthier bits of that suburbia) are more likely to oppose housing development. Well knock me over with a feather, that’s some insight!

Here in Cullingworth – which is shaded a deep red for ‘we don’t like housing development’ – the truth is more nuanced and won’t be helped by the patronising ‘communications planning’ that Shelter propose.  What we don’t want – and will die in a ditch to stop – is the sort of massive development that will change the entire nature of the place. But the development of a few houses here or there we can live with – yes we want sensitivity, we don’t like fine old houses being knocked down to make way for ‘ticky-tacky’ houses and we’d like the developer to pay attention to traffic issues and road safety.

Yet Shelter – armed with their geodemographics and an opinion survey – want to cast us as unbending NIMBYs. Not only are they wrong, they are misleading developers and undermining the real efforts of councils, planners and local communities to meet housing needs without destroying the nature of a place. Plus of course – as we found out 20 years ago – geodemographics, even supplemented with other data, is a pretty blunt instrument for marketing. A lesson Shelter (who probably didn’t talk to any direct marketers in creating their whizzo system) still needs to learn.

....

Aim! Fire! - salty thoughts on targeting

Many moons ago I was Account Planning Director at a direct marketing agency. And there we talked about targeting. Obsessively. None of that exploitative ad-gabble about 'brand equity' or 'share of mind' - we talked about targeting. About improving targeting - getting ever closer to the direct marketer's holy grail. To contacting you at the very moment you want to buy what I have to sell you.

At the heart of improved targeting is information - data about you and what you do. There's the obvious stuff - name, address, telephone number and, these days I guess, e-mail. But these are just ways to reach you - on their own these data simply allow me to contact you, they do not allow me to target. If all I have is your address my targeting is determined randomly since I do not know whether you are more or less like to want my goods of services.

So we collected other information and we developed very clever (we thought) systems based on geodemographics (the 'birds of a feather flock together' principle) and psychographics (or 'lifestyle targeting' as the salesman would put it). These systems - built on the back of the electoral register, credit referencing information and other available behaviour data - were aggregated. We didn't know the information about each individual just a set of likelihoods determined by multiple regression analysis. But, couple with a list, we were able to identify the places where the birds who liked our product were flocking, and in doing this to improve our targeting from random.

But this 'profiling' approach - for all its merits and efficiencies - doesn't work that well from the marketers perspective. Despite all the clever number crunching and the melding of more and more information, real behaviour data was always better. Let me explain. Geodemographic profiling will tell me where my customers are concentrated - but it won't tell me enough about my customer's immediate neighbour to generate a sufficient uplift in response.

So why - other than Sunday afternoon indulgence - am I burbling on about targeting? Well, it seems to me that we have to resolve the use of targeting - not by businesses but by public authorities. I recall trying to persuade Bradford Health Authority to use targeting to improve the performance of public health campaigns. Rather than scattering information far and wide in the hope that it reaches the target group, why not use thse geodemographics and other data to get the message directly to the person at risk.

Back then we were sending messages about AIDS to 75 year old grannies and I'm sure not much has changed today. Take salt. For some (but not all) people with hypertension reductions in salt intake are highly recommended as a means to manage heart attack risk. For the rest of us it really doesn't matter - our salt intake in no way constitutes risky behaviour even if it is far above so-called recommended levels. The 20% of the population for whom salt is a risk factor can be easily identified through a simple test (GPs could do this) and the rest of us could go on with having food actually tasting of something.

The same approach could be adopted with other risky behaviours - rather than spending millions sending messages to a general population that isn't at risk (for example in their drinking, smoking or drug habits) we could direct that funding towards those whose risky behaviour does present a problem. We could target but we don't. We could use geodemgraphics, medical records and much else to improve public and primary health but we don't. We opt - for reasons of 'fairness' and equity - to spend the money on general campaigns produced by grand, flash and fancy ad agencies rather than intelligent, targeted direct marketing agencies. And we prefer to ban the agent rather than address the problem user.

Targeting public health campaigns would have a number of beneficial outcomes; Firstly it would mean we get better health outcomes from the spending; secondly, it would get away from the finger-wagging nanny state approach to health campaigns; and third it would allow more risky behaviours to receive public health campaigns. And, of course, it would mean that folk like me who are not massive public health risks and know what we're doing are less pissed off by the hectoring doctors and their chose fake charities.

....