Adoption of geodemographic and ethno-cultural taxonomies for analysing Big Data

Abstract

This paper is intended to contribute to the discussion of the differential level of adoption of Big Data among research communities. Recognising the impracticality of conducting an audit across all forms and uses of Big Data, we have restricted our enquiry to one very specific form of Big Data, namely general purpose taxonomies, of which Mosaic, Acorn and Origins are examples, that rely on data from a variety of Big Data feeds. The intention of these taxonomies is to enable the records of consumers and citizens held on Big Data datasets to be coded according to type of residential neighbourhood or ethno-cultural heritage without any use of questionnaires. Based on our respective experience in the academic social sciences, in government and in the design and marketing of these taxonomies, we identify the features of these classifications which appear to render them attractive or problematic to different categories of potential user or researcher depending on how the relationship is conceived. We conclude by identifying seven classifications of user or potential user who, on account of their background, current position and future career expectations, tend to respond in different ways to the opportunity to adopt these generic systems as aids for understanding social processes.

Introduction to and structure of the paper

The purpose of this paper is to consider a currently unexplored field, namely the reasons for the uneven adoption of two new forms of generic classification, derived from Big Data, which are commonly used as general purpose, cross-industry and cross-application standards when analysing social behaviour. Many innovative methods have been developed by social scientists to analyse Big Data. This paper does not however focus on methods. Rather, it examines the adoption of geodemographic or postcode-based classifications such as Acorn and Mosaic which classify people on the basis of the demographics of the neighbourhood in which they live and Origins, an ethno-cultural classification, which infers cultural background from a person's name. ¹

The relevance of these systems in a world of Big Data is that they enable people of known identity to be categorised by proxies for social status and ethnicity quickly, inexpensively, without the need to develop a bespoke classification system and without the need for them to respond to a survey questionnaire.

In an introductory section we contextualise our discussion within the framework of the wider discussion regarding the adoption of Big Data as an alternative source of empirical evidence to that of the social survey. We then explain the origins of generic taxonomies whose standardised categories are derived from the use of Big Data and discuss their potential complementarity with better established taxonomies whose widespread use results from the ease with which they can be included as questions in survey questionnaires. It is important in this context to distinguish, on the one hand, generic taxonomies, these being designed by one group of researchers for use by an extensive population of other third party users, from bespoke taxonomies, these being designed by a user group primarily for use in its own research projects or from research projects whose purpose is aimed at improving quantitative methods for optimising different forms of classification. ² It is the use of the first category, generic taxonomies, which is the exclusive subject of this paper. The other categories are already well-documented and discussed in the academic literature.

To understand the uneven adoption of generic taxonomies among the groups of potential user for whom they have been designed we review some of the features of these new taxonomies which, on the basis of our experience, prove problematical for certain categories of user and attractive to others. This leads us to a consideration of the theoretical basis on which they are constructed and the challenges these new generic taxonomies present to existing theoretical conceptualisations.

From this consideration of their features and relationship to theoretical constructs we believe we are in a position at the conclusion of the paper to identify a number of broad categories of user or potential user and to characterise them according to how likely they are to adopt these taxonomies.

Big Data as an alternative to the social survey

First let us consider the wider context of the adoption of Big Data in the social sciences. In their seminal paper ‘The Coming Crisis of Empirical Sociology’, Savage and Burrows (2007) warn that the future effectiveness of the social science community is in danger of being prejudiced by over reliance on the social survey as a tool for generating quantitative evidence. ³ In their opinion there are many fields of interest to the social science community that could be more effectively investigated using large-scale data sets of the kind generated as by-products of the digitalisation of transaction processing systems (“Big Data”) which, generally speaking, contain data on the whole population as opposed to a sample drawn from it. The key advantage therefore that most Big Data databases hold over traditional respondent surveys files is that they contain information on a larger number of cases, the information is more up to date and is typically more granular. For many researchers a further advantage of Big Data is that it delivers information about revealed behaviour rather than stated preferences.

It is argued that by failing to broaden their methodological tool kit the practice of social scientists and policy analysts stands apart from that of commercial organizations. Almost all the largest commercial organisations now accept the strictures of management consultants such as Peppers and Rogers (2003) that a critical source of competitive advantage lies in the effectiveness with which a company uses transactional information to understand customer needs. This, they argue, can only be achieved through the creation of large data repositories. The purpose of the Savage and Burrows paper was to describe what, by comparison with the commercial sector, constitutes a comparative failure to adapt to these new opportunities. The paper did not set out to provide an explanation for this failure.

Of course Big Data will never entirely replace the findings of questionnaire-based surveys. The challenge to the social sciences is to find an appropriate balance. Though one can argue about what constitutes an appropriate balance, in 2007, the year the Savage and Burrows paper was published, there was little evidence of social scientific researchers being among the vanguard of Big Data users. The situation today has not changed markedly. The high level of interest that the paper generated has had only a moderate impact on the usage of Big Data notwithstanding the exhortations of funding bodies such as the Economic and Social Research Council (ESRC) and the contents of follow up papers such as Uprichard et al. (2009).

In our view, the explanation for the modest impact of the paper on the culture of research practice is not that the intended audience of social scientists did not hear or did not understand its message. More likely explanations lie in the inter-connection of cultural practices, institutions and assumptions which characterise the organisational structures in which different categories of potential user operate. Institutional challenges include the obstacles involved in gaining access to the typically commercially sensitive, proprietary databases controlled by privately owned companies for whom the data is a valuable asset – witnessed by competition to acquire the Tesco club card subsidiary Dunn Humby, the challenges that fragmented research communities can have in mobilising the resources to store and manipulate databases of great size and complexity and, as we explain later, the shorter time horizons over which researchers' investment in new modes of analysis need to be amortised given the length of time for which the projects or research teams in which they work are funded.

Cultural explanations (by which we mean the norms of accepted academic practice) include the current privileging of explanation over description, assumptions regarding the salience of causation, verifiability and the importance of theory and concepts. These practices are reflected, for example, in the statistical techniques which students are taught and tend to rely on, the criteria by which grant applications are evaluated and the yardsticks by which the output of institutional research is judged. A further and fundamental contrast with ‘mainstream’ social science involves the way in which the latter often proceeds through the use of the ‘comparative method’ (across for example time, space or different cultures) and the analysis of social change. In commercial applications relatively greater attention is given to evidence and less to theory and what conceptualised originate from the use of data are likely to be treated as commercially sensitive. Crudely put, the academic community relies on traditional methods as a common currency by which the richness or otherwise of its research endeavours can be evaluated on a peer basis.

Contrary to the presumption of Thrift (2005) that capitalism finds no difficulty in exploiting Big Data for commercial advantage, an analogous debate can be found in the journals of the commercial market research industry (Mouncey, 2014). Here it is asserted that mainstream market research companies, notwithstanding their expertise in how to interpret the output of social surveys and present its implications for corporate strategy, have consistently failed to apply these skills to the analysis of “Big Data”. This emerging market has therefore been left open to exploitation by specialist data analytics businesses many of whose founders' careers were built from expertise in statistics and information technology rather than in social scientific research.

When considering the potential evidence value of Big Data as compared with traditional survey sources, it is important to recognise a key dilemma faced by the researcher. That is how to reconcile the competing advantage of a rich source of behavioural evidence with a comprehensive array of key demographic classifications.

In comparison with sample surveys such as the General Household Survey (GHS), Big Data datasets tend to contain a proportionately greater wealth of behavioural data but at the cost of much less rich demographic content. This difference results not from the intrinsic characteristic of Big Data datasets. It is not necessarily that they are “big” or that they are differently structured from other forms of data, though they often are, but that they involve the secondary use of data collected intentionally or ostensibly for some other purpose. The predominant sources of this by-product data are transactional processing systems, often updated remotely, digitally and in real time in which there is no baseline data or anchoring in categories of socio economic status (SES) such as occupation, ethnicity and marital status. Such systems might be ones that process customer orders, maintain the financial balances in customer accounts or enable digital connectivity, for instance through delivering access to websites, engaging with social media, enabling telephone traffic or providing information on spatial location. Increasingly the data is also collected intentionally from applications whose primary use provides “permission” for its secondary use. Either way, they arguably generate much richer sources of behavioural data than survey questionnaires. However given that an organisation's customer/client insight team does not by definition have ultimate control over what information is contained within a Big Data database, analysts are restricted to accessing whatever data fields were specified as necessary for the efficient processing of the transactional system. In most cases these fields do not include measures of class, age, ethnicity or gender.

For the academic social scientist the absence of these fields is likely to constitute a major obstacle since, as a general rule, these classifications are the basis on which much social scientific theory and hence explanation has been developed. For analysts in commercial organisations this is much less problematic since the theoretical concepts on the basis of which businesses make decisions are to a much greater extent developed around the relationship of one behaviour to another, not the relationship of behaviour to demographics. Whereas academic researchers may be more likely to use constructs in the form of general rules, commercial researchers tend to use constructs which are specific to individual industrial sectors or even individual businesses. Demographic details such as age and gender may of course be collected via a parallel system such as the application processing system for a loyalty card. Where this is not the case social scientific research usually has an interest in finding a method of inferring them from other data fields that are available or in using proxies for them.

This is where the new generic taxonomies play a particularly useful role and why the increasing use of Big Data has resulted in their growing use as alternatives to traditional taxonomies. Measures of neighbourhood type and cultural background can be appended to any Big Data dataset that contains a customer/client name and address. This can be done quickly, cheaply, retrospectively or in real time, unobtrusively and without the distorting effects of high levels of non-response (Harris et al., 2005; Sleight, 2004).

Taxonomies enabled by Big Data

Accepting that the proprietary nature of its content constitutes a barrier that restricts access to Big Data datasets by the wider research community, in this section of the paper we consider it useful ⁴ to examine the cultural and institutional constraints on the adoption of Big Data within the more restricted frame of new generic taxonomies derived from Big Data. In particular, we consider the origins, data sources and methods used to build these classifications, the commercial organisations that deliver them and the role they play in the analysis of Big Data sets. This particular segment of the world of Big Data is referred to, but not explicitly differentiated, by Savage and Burrows (2007). It is of particular relevance to the wider discussion on the adoption of Big Data since in this particular segment neither cost nor proprietary ownership constitutes a significant barrier to use. We define this segment as the relatively new generic taxonomies such as the geodemographic classification systems Acorn and Mosaic and the ethno-cultural classification system Origins.

Mosaic and Acorn are taxonomies which enable customers/clients/citizens to be classified into some 60 or so standard categories according to the type of residential neighbourhood in which they live. Origins is a taxonomy which uses personal and family names to infer a person's most likely ethno-cultural heritage. This classification system uses a complex set of rules which place individuals into the most appropriate one of some 200 different categories.

Whilst very early versions of Acorn and Mosaic were constructed solely using census statistics, the versions of these taxonomies in current use have been constructed using access to a variety of Big Data sources. Their value to commercial and social research is that from the name and postcode field on a Big Data dataset it is possible to append both neighbourhood type and inferred heritage. Using these taxonomies as the axes in tabulations makes it possible to analyse behavioural data on different datasets using a consistent basis for comparison. A perceived benefit of these taxonomies to the social sciences is that linked to Big Data they go some way to bridging the ‘quantitative/qualitative’ distinction by enabling access to a richness of interpretative material that is often beyond the capability of conventional statistical analysis. Few investigations of social behaviour take place without reference to a taxonomy derived from one or more of the categories by which social class, race, religion, age, gender, geography and household composition are commonly represented. ⁵ It is the analysis of behaviour against the categories contained in these taxonomies that provides the evidence against which theories can be tested and developed. Inevitably there is a tendency for taxonomic categories of this sort to become institutionalised in the form of “standard” classification systems. Office of National Statistics (ONS) standard socio-economic classes and ethnic groupings are examples. Consistency of categorisation makes it is possible to compare cross-tabulations sourced from independently collected behavioural databases.

Big Data poses a challenge to such taxonomies and they to it. It is easy to overlook the fact that the practices by which people are categorised according to their social class and ethnic origin were developed in an environment where the survey was the principal method of collecting behavioural information. ⁶ To practitioners familiar only with the use of social surveys as a source of quantitative evidence it is not necessarily apparent that such taxonomies are often not practical or, if they can be obtained, not necessarily optimal in a Big Data environment. Likewise the proxies for class and ethnicity which it is possible to apply in a Big Data environment are not necessarily characteristics which users of survey questionnaires would find it easy to adopt without an investment of time and concentration, both of which may be in short supply.

Taxonomies designed for use in survey research, whilst designed to be predictive of differences in behaviour, have necessarily been designed in a form which caters for the practicalities of the questionnaire as a data gathering instrument. Respondents tick the box which matches their age band and gender and, from the occupation and the country of origin that they write in, the survey administrator uses look-up tables to translate occupations into social classes and countries of origin into standard ONS ethnic groupings. Multi-variate taxonomies such as Mosaic, Acorn and Origins work in a quite different way. Given that Big Data repositories are rich on behavioural information but light on conventional demographics, they are more likely to employ complex grouping algorithms, based on fuzzy logic and implemented using various forms of cluster analysis. Customers / citizens are then organized into “natural” classes on the basis of multiple attributes crossing a number of different domains.

The sense in which these classes are “natural” (as opposed to “purposive”) is that the categories are deliberately not created in such a way as to reflect pre-existing theoretical representations of social structure. They are what most social scientists would therefore characterise as ‘empiricist constructs’. It would be odd if there were no correspondence between their natural clusters and the categories of traditional taxonomies but ultimately such correspondence is incidental rather than designed. Since most social scientific theory is based on constructs of which conventional taxonomies are designed to be the representations, a frequent criticism of “commercial” sociology by academic social scientists is that it uses concepts that are a-theoretical, ⁷ that they have little more than descriptive value and that they have little to offer by way of causal explanation. To this consumer marketing analysts reply that the new generic taxonomies provide a more effective source of causal explanation than single dimensional constructs that are inherently difficult to link to behavioural information and that the apparent lack of development of causal theory in commercial organisations cannot be presumed solely from the absence of papers authored by commercial practitioners in academic journals. Another key distinguishing attribute of taxonomies arising from “Big Data” is that they are multi-variate by contrast with most (though not all ⁸ ) taxonomies against which survey behaviour is tabulated. On grounds of practicality these tend to be built from a single criterion as respondents necessarily find it difficult to operate a complex rule set involving multiple criteria. ⁹

Where different fields of a questionnaire are combined, as for example with indices of deprivation, the derived taxonomy is typically created by summing scores on data variables each of which acts as a surrogate for a more general and less easily defined syndrome, deprivation. This contrasts with the practice of defining categories using permutations of values from variables from different domains (such as families with children without a car). ¹⁰ Thus one would expect a clustering of retail loyalty card-holders based on previous purchasing history to produce a set of natural clusters which can be interpreted partly but not exclusively in terms of age, income and family composition. One would also expect the categories to be differentiated in terms of take up of promotional offers, use of own-label versus manufacturer brands and frequency of shop visits. The manner in which these clusters are described would align more closely with the language used by qualitative researchers (or even political strategists) than with the language of professional demographers.

A common application of Big Data is to build bespoke taxonomies which segment an organization's own customers or clients on the basis of transactional information. In this instance, it may be less important for the categories to be comparable with widely used classifications than would be the case in an academic social scientific or public sector application; the organization concerned may place little value on the added theoretical understanding of the sort that would be obtained by aligning the individual classes with professionally accepted standards. Nevertheless there are contexts in which commercial organizations benefit from using generic taxonomies recognised as standard by the syndicated industry databases that they subscribe to, such as the National Readership Survey and the British Market Research Bureau's Target Group Index.

The complementary relationship between user specific and generic industry-accepted taxonomies is nicely illustrated by the behaviour of retail banks many of whom place customers into the categories resulting from the bespoke taxonomies they build using previous customer behaviour. In such a taxonomy it is likely that at least one cluster will be characterised by customers who have tended to prove unresponsive to the promotion of new financial products. When this bespoke taxonomy is further segmented on the basis of the categories of a generic taxonomy the bank's insight team is now likely to be able to distinguish between those relatively inactive customers who tend not to respond to promotion because they have little disposable income from those who are unresponsive because they undertake the majority of their financial transactions with a competitor or a specialist adviser.

For all these reasons, it is unlikely that a generic taxonomy such as Mosaic, Acorn or Origins will precisely replicate any of the measures on the basis of which social scientific theory has been constructed and it is not logical to evaluate their validity solely on the basis of how accurately they replicate a traditionally asked demographic construct. Their categories are examples of different constructs, seldom as clearly articulated as those developed in academia, and are designed to be used in different contexts. These constructs have been slower to find their place in academic social scientific than among media and political classes.

Clearly there are contexts where it is too inappropriate or too expensive to obtain from each customer the information needed to place them in a formal taxonomy of the sort that is used in a survey questionnaire. A charity would almost certainly experience a drop in the value of its donations if it asked potential donors for their annual income. A chain of betting shops would receive short shrift from punters if they were asked their religion or ethnic identity. In contexts such as these it is more practical for an organization to make use of one of the Big Data derived classifications that are marketed by information service providers as alternatives to class and ethnic origin.

Linking an externally built classification system to one's own customer records is possible only if the two have access to a common match key. In contexts where goods have to be delivered or bills posted no-one objects to providing a supplier with a record of their name and address. ¹¹ As a result the various components of a person's name and address are the fields on customer files which are most often used as match keys used for appending generic classifications.

The various elements of a person's name and address provide a surprisingly rich range of material on the basis of which inferences can be made. Many women's marital status can be inferred from whether they style themselves as Mrs, Miss or Ms. Gender can be predicted with a 90% or greater accuracy from the personal name. Particular personal names, such as Ivy and Wayne, are associated with particular generations and socio-economic backgrounds. Personal and family names, such as Mohammed or Van Dijk, often indicate ethnic origin, language, religion or denomination.

The postcode field provides perhaps the most valuable source of information since, even in the era of social media, the social influences to which people are likely to be exposed are typically a function of the demographics of the immediate neighbourhood in which they and their families live. All these inferences are examples of the potential value of by-product data generated as a result of the operational requirement to capture name and address for the purposes of delivery and/or billing.

Names and postcodes are the two elements of the name and address which most lend themselves to the construction of serviceable taxonomies. Postcodes are the units of geography now used by Experian, CACI and other vendors of geodemographic classifications. ¹² These classifications make use of over 200 different measurements, derived from official, administrative and census sources, resulting in each postcode being assigned to the most appropriate one of 60 or more geodemographic categories. By analogy Origins is a system which associates each record contained in an organization's name and address file with one of 200 ethno-cultural backgrounds. A Big Data database containing the personal and family names of 600 million different adults was used to build this classification system.

When used in combination with the software that appends them to customer / citizen files, taxonomies of this sort offer a variety of attractions. They provide a very fast and inexpensive method of categorising very large numbers of customer or citizen records. They are non-intrusive. Their match rate of over 99% means that the validity of results is not compromised by uneven response rates. ¹³ The categorisation is consistent, can be updated on a regular basis and is independent of the context or the operation of the classification process. It permits consistent comparison between analyses of different datasets and at different times and, in respect of ethnicity, a consistent set of categories irrespective of the countries where the names have been collected. ¹⁴

It is thought that postcode classifications are used by the majority of UK ¹⁵ organizations with databases containing information on two million or more customer address records. Because geodemographic classifications have been appended to respondent data for each of the most widely used market research surveys researchers can examine their predictive power across many thousands of behaviours. On balance the level of this discriminatory power is broadly similar to that of standard measures of social grade. Qualitative fieldworkers such as Butler and Robson (2003) have reported a very close correspondence between their own impressions of the character of communities where they have conducted research and the portraits used by geodemographic classifications to describe postcodes within their study areas. Indeed given the granularity of the postcode system, with over 1.4 million geographical pieces, and the 60 or so different classifications that a taxonomy usually identifies, it would be surprising if a natural taxonomy of neighbourhoods based on the resources of Big Data could not provide a much more fine grained and accurate description of a neighbourhood than any metric based on a single-dimensional domain.

To a large degree, this results from the freedom of computer algorithms to create natural clusters which are unconstrained by constructs derived from pre-existing theory and from the use of a single theoretical dimension. The Mosaic label “Liberal Opinion” is a good example. It was coined to describe neighbourhoods characterised by, among other things: large numbers of recent graduates, particularly in the humanities; living in privately rented purpose-built or converted flats; low levels car ownership; low child populations, especially among school age cohorts; many children born to mothers in their 30s.

No single demographic dimension would be capable on its own of identifying this type of neighbourhood, or what, in Bourdieu's terminology, can be described as a “milieu”. It is our contention that academic social science and commercial sociology can come together fruitfully when focusing on emergent social groupings such as this. New generic classifications built from Big Data should be able to provide greater nuance to terms such as ‘new urban middle class’ or even ‘gentrifier’ as well as locate them in space and enable their purchasing preferences to be recorded.

That a category such as “Liberal Opinion” does have inherent validity is evidenced by the results of the analyses of market research surveys to which Mosaic codes have been appended. These identify diagnostic behaviours such as frequent foreign holidays especially to offbeat locations, early adoption of new personal digital technologies, extensive use of restaurants and coffee shops, heavy readership of newspapers and on-line media channels, frequent visits to cinemas, theatres and exhibitions and support for environmental and international charities. In the 2010 General Election Conservative candidates achieved consistently lower than average swings in constituencies with the highest proportion of electors in this type of neighbourhood. The downside of using such a taxonomy is that the clusters are re-optimised at least every time the results of a new census are published. This makes it particularly difficult to undertake temporal comparison.

The ability of a taxonomy based on multi-variate criteria to create more meaningful categories than one based on a single survey question is equally relevant to the classification of people by ethno-cultural background. Clearly there needs to be some method of recording a person's heritage. But is that best obtained by asking them their religion, their country of birth, their ethnicity, their self-identification, the country where their parents were born or when they entered the UK? Each of these questions has appeared on a post war British census but we contend that no single question on its own provides an optimal basis for a general purpose taxonomy which will maximise discrimination when comparing the behaviours of members of minority communities.

Origins operates at a variety of levels of granularity. At the coarse level the natural categories are interpreted from the overlap of race (or colour), religion and geography. The natural grouping “Muslim” includes Black Africans with predominantly Muslim names. Detailed categories such as Fijians and people of Indian descent from Surinam would be placed in the same general grouping as people with “Hindu or Sikh” names who are geographically South Asian. East Asian is a group primarily distinguished on the basis of geographical location whilst “White” and “Black” are distinguished primarily on the basis of appearance.

When the Origins classification was built the intention was not to create a taxonomy incorporating religion and language as well as race. However the optimisation algorithm used to build the taxonomy created a number of categories which could be effectively interpreted only by using all three of these dimensions.

Features of new taxonomies that influence potential users

In order to understand differences in the speed of adoption of these new generic taxonomies it is important to undertake a brief summary of some of their key features which differentiate them from traditional classifications and which are relevant to potential users.

For some potential users these differences present no hindrance to adoption. Indeed for some applications and user categories these differences represent clear advantages. For other potential user groups these features may be off-putting.

When business planners in commercial organisations such as Experian or CACI evaluate the market potential for a new segmentation system their evaluations would typically take into account an inventory of potential applications which are then assessed in terms of potential impact and hence potential return through sales. At this stage in the product cycle these projections are likely to be based on an assessment of potential benefits to different user groups, it being difficult to gauge the extent to which factors (including those which we have broadly termed institutional and cultural earlier in the paper) are likely to influence the rate and level of adoption in different areas of application.

In due course the deployment of the sales team is adjusted on the basis of where it is easiest to win new business. Any variance between the sectors of anticipated and actual usage is typically explained by the management of the sales force in terms of the rationality or irrationality of actors or even their levels of intelligence and insight. Uneven adoption is not interpreted in terms of the differences of perceptions and expectations that result from differences in culture and organisational structure.

In the following section we set out the key features of these classifications which contribute to differential rates of adoption among different potential user groups. These features are discussed under seven headings as follows.

Actionability

The characteristic commonly used to justify the adoption of geodemographic classification in the commercial sector is its “actionability”. Whilst it was perfectly practical for a media buyer to select advertising slots on commercial television channels on the basis of whether the social class and age of a programme's audience matched that of a brand's users, social class and age could not be used to target contact channels such as direct mail, door to door distribution, tele-marketing or customer mailings which, by their very nature, have much greater potential for reaching finely-grained audience segments.

For these channels, taxonomies based on names and addresses proved more “actionable” since these communications channels could be targeted right down to the person, household or street (postcode) level. The use of such communication channels is not confined to commercial organizations. Charities, political parties (Webber, 2006) and a number of government organizations regularly use these channels to communicate with citizens and Big Data based taxonomies have been proved as popular with them as with the marketing departments of commercial organizations.

Maximising the precision of targeting is a more pressing objective for people involved in “direct” marketing and the use of generic geodemographic taxonomies for behavioural analysis tends to be restricted to the identification of the market segments that generate most profitable response to communications.

Perceived purposes

People who work in academic and not for profit organizations can harbour feelings towards the taxonomies of commercial organizations that range from mild scepticism to ideological hostility – and of course this is a two way street with commercial sociologists declaiming their academic colleagues as not living in the same world. Common objections are that the construction of these products is not methodologically rigorous, that build processes are not transparent and that the purposes for which they were designed are not aligned with the requirements of social research. Many of those who have been exposed to generic taxonomies built from Big Data are unaware that the methods used to build geodemographic and ethno-cultural taxonomies originated within the public sphere.

The methodology currently used to build geodemographic classifications is largely based on that developed during the Liverpool inner city deprivation study sponsored by the Department of the Environment, Planning Research Applications Group (PRAG) (1975), the objective of which was to differentiate neighbourhoods on the basis of their different manifestations of deprivation. Cummins et al. (1999) explain how tables that infer ethnicity from names were first built by officers of Bradford City Council with the aim of improving access to public service by people of South Asian origin.

The concern that the methods are unacceptably lacking in transparency undoubtedly results in part from the difficulty academic social scientists have in accessing the journals in which papers (such as Webber, 2000) which describe the methods used to build them have been published.

These are not factors which generally weigh heavily on the minds of commercial users or the brokers between academia, politics and the media but they are powerful influences in some sections of the academy although the primary sense of ‘unease’ is the task, generally ignored by the developers, of interpreting the “natural” clusters arising from the classification process within the pre-existing theoretical constructs familiar to social scientists. A secondary concern is the difficulty of undertaking time series comparisons given that the taxonomies change every 10 years. Some of these concerns are expanded on in the following section.

Evaluation of methodology through independent replication

It is often held that taxonomies based on Big Data are not susceptible to independent replication. The ability to replicate findings based on free access to data sets is a central tenet of research in the natural sciences where it is arguably easier to control for external conditions than it is in the social sciences. In the social sciences it is reasonable to suppose that replicability is more relevant where an objective of much research work is to demonstrate causal connections or where mathematical models are used to generate future scenarios.

Where geodemographics are used for description, replicability is less relevant. As a rule differences in the methodologies used to build classifications have some but very little impact on the extent to which different classifications can predict variations in behaviour. ¹⁶

The concern with verification cannot be considered other than in the context of a belief in whether or not there is a truthful representation which it is the objective of research methods to try to replicate. If such a belief is not held any solution is necessarily viewed as the product of subjective assessment.

Commercial organizations, which are primarily concerned with the ability of technology to deliver performance objectives, tend to focus less energy on methodological verification than on tests to determine whether a new taxonomy will outperform those taxonomies currently used to support key business processes. Confronted with such a classification for the first time commercial users like to enter their home postcodes and check whether the geodemographic portrait matches their impression of their neighbourhood. Researchers are more likely to want to ask questions about methods and motivations although anecdotally many academic researchers do precisely the same in terms of entering their home postcodes and those of close friends. For the commercial user the reputation of the product and of its vendor are of paramount influence. The quality of documentation, support and ancillary consultancy are also critical buying considerations.

Definition or description

An inherent problem with a taxonomy which is based on multiple domains is how best to label its categories. This is particularly an issue for natural rather than a bespoke classification. Tensions necessarily exist between whether labels should aim for a precise definition or a memorable and insightful description.

A recurring complaint of some members of the academic research community is the subjective and imprecise character of the labels used. Quantitative researchers, in particular, tend to dislike the way in which Mosaic and Acorn labels limit themselves to 20 characters. This limitation is often assumed to follow from a desire to satisfy the requirements of commercial users. Such labels tend to resonate better with qualitative or mixed mode researchers. By contrast commercial users prefer to use labels such as “Global Connections” which are descriptive, easy to understand and to remember and not overly precise. Their use comes more naturally to those who work in advertising agencies and the media who are accustomed to the need to distil the findings of focus groups into tag lines such as “Worcester Woman” or “White Van Territory” which are understandable to their clients.

Whilst clearly it is possible to define categories in formal classifications such as “persons aged 45–64” or “South Asians” in few characters, it is impossible to create succinct definitions of complex multi-dimensional categorisations. Indeed multi-dimensional classifications by their very nature challenge the disposition of those for whom exactitude of measurement may be a more important consideration than deeper understanding of a complex world.

For the commercial user an important consideration is to be able to identify and attach labels to groups such as “Caring Professionals” or “Coronation Street” which are often subjectively felt to exist but which are otherwise difficult to define and for which specific modes of communication are believed to be appropriate.

For qualitative academic researchers the appeal of the finer grained categories produced by a natural grouping is that they may be able to sharpen the spatial and theoretical definition of concepts already current in social research. To them labels such as “New Urban Colonists” or “Summer Playgrounds” are not just target groups that can be reached on the ground but physical conceptualisation of social processes which, because they are difficult to represent in the form of a survey questionnaire, are difficult to pin down in practice. They offer an opportunity to overcome a dilemma which many academic social scientists struggle with, that of integrating the social and the spatial.

A good example is the choice of the Mosaic label “Alpha Territory” for the name of an ESRC funded research project which investigates the impact of London's increasingly global population of supposedly footloose, ultra-high net worth residents on patterns of association among elite London neighbourhoods (Burrows, 2013). It could legitimately be complained that the term “Alpha” has no formal meaning is social research. Nevertheless it is an example of a term that is both economical and efficient in denoting a group of very high achievers who are strongly oriented towards delivery, whose culture is dismissive of anything that is sub-standard in quality or performance and whose language regularly borrows from the terminology of global capital.

Multi-dimensionality

The persisting popularity of indices of deprivation and of school league tables suggests the preference of the UK political elite for one-dimensional ranking over multi-dimensional classification. The predilection for good–bad rankings is even more evident in the pages of the British Sunday papers. In this context it is ironic that the original justification for the development of geodemographic classifications was to identify the different types of deprived neighbourhood for which different policy interventions would be appropriate. This was intended to replace a policy under which all priority area funding was channelled to the same “worst” areas, irrespective of whether or not the programmes were appropriate to their needs. ¹⁷

As a general rule marketers, whether in the commercial or public sector, are much more focused on segmentation than academic social scientists, policy people and journalists. The aim of marketers is to separate customers into discrete groupings on the basis of differences in needs and motivations and then improve efficiency by tailoring communications to the needs of each market segment. This approach to customer relationship management is reflected in the form of the metrics organizations want to be able to access on their databases and which they use to track the effectiveness of their campaigns.

Whether or not this communication process is undertaken by commercial or not for profit organizations is less important than people may suppose. One of the earliest types of organization to adopt geodemographic classification were US and UK political parties. Whilst many constituency MPs wax lyrical about the contribution these systems can make to their re-election, it seldom occurs to them that the very same tools could be equally effective in improving the targeting of public services communications to their constituents.

Systematic bias

A key consideration of any taxonomy is whether its classes contain any systematic bias in relation to behavioural variables against which they may be analysed. This is particularly relevant in two respects. Where errors are random it is usually safe to use a taxonomy for measuring and comparing distributions of a population by classes of that taxonomy. It also suggests that the use of classes within a taxonomy to model behavioural outcomes is likely to be reasonably accurate. For example, it is statistically accepted that different social classes have different propensities to vote Conservative or Labour in a UK general election. Let us suppose that these different propensities were uniform throughout towns and cities of the United Kingdom. In this case we might reasonably suppose that it would be possible to predict the Conservative and Labour shares of the vote in each parliamentary constituency simply by multiplying the proportions of its electorate in each social class by the national propensity of each social class to vote Conservative or Labour.

It might surprise many to learn that on the basis of this arithmetic the Conservative party would not have won a single one of the 165 seats it won in the UK general election of 1997. The reason is that in upper class neighbourhoods, upper class electors are more likely to vote Conservative than where they are in a minority and vice versa. The social class of your neighbours provides as good an ability to predict how you will vote as does your own social class and one which is more uniform in its predictive power in different types of constituency. Likewise it has been shown by Webber (1975) that although the level of unemployment in Liverpool in the 1970s was far lower among workers who were professionals or managers than among unskilled workers, a professional or manager living in a neighbourhood of inner city council flats was more likely to be unemployed than an unskilled worker living in a prosperous suburban neighbourhood.

These are just two examples of individual behaviour being subject to neighbourhood effects, replicated in countless other studies – in other words, the behaviour of immediate neighbours affects the behaviour of the individual. In these two cases it would appear either that what we define as a “Professional or Manager” conceals multiple different sub-populations or otherwise that professionals and managers behave in systematically different ways according to where they live. There is systematic bias, in this case by type of neighbourhood.

Logically it would be reasonable to suppose that the categories created by a natural classification, which is constructed using a wide array of different input variables and where no single one dominates the clustering algorithm, will be much less subject to this sort of systematic bias than traditional classifications which are defined using the response to a single survey question. This goes some way to explain one of the attractions of neighbourhood classifications to retail multiples who need a reliable method of estimating likely levels of demand for different products and services in the catchment areas of locations where they might open a new store. Since the propensity of different geodemographic types to consume different products is readily available from syndicated survey sources and since the distribution of Mosaic types within particular catchment areas is also known, a basic model which multiplies purchasing propensities in each Mosaic type by the proportion of the catchment area that they represent will provide an estimation of local levels of product demand relatively free from systematic error.

Self-identification

Comparing taxonomies arising from Big Data and the formal classifications designed for use in survey questionnaires throws up interesting tensions regarding the role of self-identification and ascription in the categorisation process. A significant number of academic commentators express mild concern at the practice of assigning an ethno-cultural classification to subjects in a database thereby denying them the opportunity (or indeed right) to specify their own self-identity. This raises a number of interesting questions. If it is right for the individual to select the ethnic origin that he or she identifies with, should not that right apply in the case of social class? And should we define areas of multiple deprivation on the basis of asking people whether they feel they or their near neighbours are deprived?

In contexts where either class or ethnicity is used as a basis for analysing citizen attitudes and behaviours to which identity is relevant, such as for example self-worth or perceptions of the degree of racial discrimination within an organisation, this would be a powerful argument. However this argument has less power when one is considering behaviours which are more likely to result from inherited characteristics, such as for example the political party a person supports. In these circumstances a more effective predictor, if it could be constructed, might be the ethno-cultural category to which a person's parents belonged.

Personal experience has taught us that asking the supermarket check-out operator or the waiter “what are you / where do you come from?” seldom elicits a favourable response. It is clearly impertinent. However it is our experience that posing the question “where does your name come from?” seldom elicits a negative response. Their name, though seemingly a very personal characteristic, is seen by most people to have a sufficient independence from their identity that its origins can be discussed without trespassing on personally sensitive territory. Indeed discussion of the origins of the owner's name will often lead to discussion of quite personal areas of their life history. This distinction is relevant when considering the high rate of non-response to ethnic questions by survey respondents as well as to the acceptability of using names as proxies.

A common response of people when asked whether the portrait of the type of neighbourhood their postcode falls matches their self-perception is that it provides only middling accuracy but, however, it is a very accurate description of their neighbours. None of us particularly like being put in a box notwithstanding the interest we all have in finding out which box others belong to.

Challenges to existing conceptualisations

In previous sections we have described the features of taxonomies derived from Big Data which can be aligned or misaligned with the requirements and cultures of different groups of potential users. In this section we examine two other problematical characteristics of new taxonomies based on Big Data, particularly those which can prove difficult when incorporated within the traditional conceptualisations which underpin social scientific theory.

(a) Timescales of inherited characteristics

When formal and Big Data derived taxonomies produce two apparently inconsistent classifications of the same individual, many researchers understandably presume that any discrepancy must be due to error of some kind. In practice these inconsistencies can often be explained by the period of time over which the influence of inheritance decays.

According to the results of the 2011 census few large English cities have a smaller proportion of people who declared themselves as Irish than Middlesbrough. However Middlesbrough is close to top on the ranking of English cities by proportion of inhabitants with Irish surnames. The reason for this paradox is that family names indicate distant heritage, in the case of Middlesbrough a heritage some four or five generations back, whereas self-identity, as recorded in the census, is a more forward-looking definition, often corresponding to the culture the respondent aspires to belong to or with which she or he identifies.

The extent to which it is distant ascribed inheritance or current self-identification that is more predictive of behaviour will vary from behaviour to behaviour. An example of distant heritage being more predictive of attitudes than self-identification is the source of support for an independent Scotland. Despite 10% of Scottish residents bearing Irish Catholic surnames the proportion who self-identify as Irish is just 1%. YouGov survey results show the 10% of Scottish residents with Irish Catholic surnames to be more supportive of Scottish independence than people with Scottish surnames by 10 full percentage points (Demos, 2014).

This example nicely illustrates a number of issues relevant to the conceptualisation of ethnic or religious classification. First, as has been long known to US politicians, that historically distant heritage can continue to have a very real impact on present behaviour. In other words the influence of heritage on behaviour can decay very slowly. The second is that the number of respondents who self-identify with their distant heritage may be far smaller than the number whose behaviour is influenced by it, in the case of the Scots Irish 10 times. Taking into account these two factors it is evident that the group of people who claim a particular identity, in this case Scottish, can easily conceal very important behavioural distinctions, in this case to support an Independent Scotland (Phillips, 2014).

Ethno-cultural background does not correspond exactly with self-identity. The two represent different components of ethnicity.

(b) The use of proxies

Given their entrenched use and institutional acceptance, it is understandable that what were originally intended as indicative proxies can over time become mistaken for the entities that they were originally supposed to be indicative of. It is easy for analysts as well as researchers to slide into the assumption that there is a 1:1 correspondence between an “indicator” and the characteristic which it was originally intended to “indicate”.

Class, which is a multi-faceted and nebulous but nevertheless a real characteristic (because we know it discriminates on behaviour), eventually becomes confused with the occupational classification which was originally intended as a proxy for it. This explains why class has fallen into so much disfavour in both commercial and academic sociology – a rare achievement. Research originators as well as research users slide into a reification where metrics are no longer separate indicators from the complex, multi-faceted concepts for which they were intended to act as proxies.

In such an environment it becomes easy to suppose that any discrepancy between a challenging taxonomy and the incumbent proxy is misinterpreted as a source of error – the incumbent has become the reality.

This slide is well illustrated by the use of an index of multiple deprivation which, as its name implies, can do no more than act as a proxy for whatever deprivations local residents experience. Once deprivation becomes unambiguously identified with its manifestations, local councillors speak proudly about their wards being in the 10% most deprived in the country when in reality they are in the highest 10% of wards as measured by a bundle of proxies for deprivation.

This confusion may seem immaterial until we find as perspicacious authority of the condition of minority communities as David Goodhart (2013) argue that a disadvantage experienced by minority ethnic populations is that they are more likely than whites to live in areas of deprivation when, in practice, this will continue to be the case by definition so long as the size of the minority population remains one of the indicators used in calculating the index.

A classification of potential users

In previous sections we have sought to explain differential adoption of new taxonomies on the basis of the features that differentiate them whether positively or negatively from traditional taxonomies. In this section we believe that it is possible to identify seven broadly different characterisations of user or potential user, each differing in terms of their likely response to the opportunity to make use of generic taxonomies of the sort described in this paper.

Had it been our intention to undertake such distinctions based on an academically rigorous and verifiable methodology this would not be possible. Instead we have the opportunity to draw on the experience of the lead author (Webber) who has worked with geodemographic data for 40 years, devising, promoting and consulting on the use of Acorn, Mosaic and Origins by diverse potential user groups. Our identification of these types also follows in part from the consideration of some of the institutional practices and cultural preferences described earlier in this paper. ¹⁸

(a) Corporate Information Managers

One important group of actors is commonly found in large, consumer facing corporates such as banks, utilities and retailers who sustain relationships with their customers via some form of customer account. These can be described as Corporate Information Managers. This group also characterises the people who manage the campaign databases used by political parties. This group would typically style themselves as information or insight specialists whose expertise lies not so much in information processing as in information management.

Much of the focus of the work of members of this group involves testing strategies which can deliver improvements in operational performance. Though many will have graduated with a degree in a quantitative discipline, few would identify themselves primarily as social scientists. That is not to say that they do not have a deep interest in citizen' / consumer behaviour. It is just that this understanding, if it is conceptualised, tends not to use the terminology commonly used among members of the academic social science community.

Though members of this group share a belief in a common set of business precepts, conceptualisations of customer motivation and behaviour are necessarily specific to the particular industrial sector within which they operate. They aspire to and often succeed in developing a holistic, multi-dimensional view of consumers which enables them to communicate in a quite nuanced manner using one-to-one communications channels supported by sophisticated campaign management software.

This group is typically early adopters of taxonomies arising from Big Data on account of the size of the databases that they manage, the ease of adding taxonomies to them and their focus on actionability. Though this group may need to apply for permission to make major financial investments in new data they are seldom required to explain to people outside their immediate team the data and methodologies that they use.

Their role is deeply embedded in the organization. Stability of purpose enables people to adopt a long term perspective.

(b) Strategic Marketing Advisers

Distinct from this group are Strategic Marketing Advisors, many of whom work in advertising agencies, brand consultancies and communications agencies. This group of actors is highly focused on the use of taxonomies of consumers, both to identify key audiences for communications programmes and to match the message to the audience. Traditionally this group has ascribed a premium to qualitative description and has focused more heavily than the previous group on understanding attitudes, values and motivations rather than just propensities to purchase specific products. This group relies heavily on conventional taxonomies such as class, age and education and is open to the use of attitudinal or multi-dimensional categories, such as the “Worcester Woman” and “White Van Man”, favoured by political strategists.

The particular interest that this group shows in taxonomies arising from Big Data is the ability of the labels to succinctly describe categories that they consider to be relevant to their clients. This group does not confuse proxies with reality. When using a taxonomy it is less concerned than Corporate Information Managers about discriminatory power or actionability than by whether the taxonomy will be accepted among the diverse sets of actors by whom they are paid. Does the taxonomy have a good “brand” reputation and are the categories instantly intelligible? Another important characteristic of this group is that most of the projects that its members work on have limited duration and their involvement with clients, being strategic, tends not to extend to the implementation of taxonomies within operational systems.

By contrast comparison with the Corporate Information Managers, this group is more likely to value Big Data based taxonomies for defining the broad segments for the purpose of brand positioning and campaign strategy. They tend not to get involved in the systems which drive one-to-one communications.

(c) Early Career Researchers

Within the academic research community we can also identify a diverse set of actors most of whom are little exposed to these forms of classification. For Early Career Researchers, there is potentially a penalty for using new taxonomies especially those that do not have self-evident correspondence to the concepts used in established theories and the metrics by which they are represented.

Their thinking tends to be less bound by the iterative process of testing and feedback that characterises commercial consumer insight specialists and, perhaps due to their lower levels of direct contact with the subjects of their research, it is easier for them to confuse metrics with the realities these metrics were originally intended to be proxies for.

Funding through research grants is dependent on the judgement usually of older and more experienced specialists many of whom have institutionalised interest in the use of existing taxonomies, not least as a result of their use in the research outputs that have contributed to current theoretical knowledge.

Research outputs are constrained by the need to communicate findings according to particular modes of expression and influence is necessarily dependent on the use of terms whose meaning is clearly understood. Thus the adoption of alternative classifications does constitute a significant investment and risk on the part of the research practitioner. The risk is that they may not have sufficient knowledge to defend their research methodologies from the criticism of methodological experts, particularly where literature reviews are likely to be thin and finding citations of acceptable rigour will be hard. The upside is that their approach can be seen by their more senior peers as ‘cutting edge’ and the associations with the private sector are increasingly regarded as ‘brownie points’ by senior managers.

Despite these drawbacks isolated individual researchers have pioneered the use of these new taxonomies in a variety of fields. ¹⁹

(d) Researchers favouring quantitative methods

Within the academic community there are small but important sub-groups. One of these, Quantitative Methodologists, are actors whose principal interest is not so much in the behavioural patterns that are the traditional object of social scientific study but in developing and critiquing statistical methods in their own right. Whilst the conduct of this form of research clearly has potential advantages in terms of methodological innovation, it is often said that with creating effective taxonomies being as much an art as a science competing methodologies are difficult to critically evaluate in the absence of direct experience of the different applications to which a taxonomy is likely to be put.

(e) Researchers favouring mixed mode approaches

Another important sub-group is Mixed Mode Methodologists, often relatively experienced researchers, who combine an interest in quantitative methods with highly specific and localised in-depth qualitative interviews. Geodemographic taxonomies are received particularly favourably by this group since they give broader geographical context to the results of localised research, provide evidence of whether particular findings are part of a more widespread pattern and help identify other communities where the processes that they have identified are likely to be most apparent.

This group is often more comfortable generating research outputs which are descriptive. The language by which they conceptualise processes often conforms less precisely to the language of established theory and the categories by which it is otherwise represented. Such researchers often have to be bilingual in terms of facing two rather different constituencies at the same time.

(f) Opinion Shapers

The network of actors that mediates between the academic community, the media, think tanks, political advisers and civil servants, Opinion Shapers have a more ambivalent attitude towards taxonomies arising from Big Data. Whilst these actors derive their reputation from disseminating ideas which are new, their influence is constrained by the fact that they are operating in a world where the recipients of their influence are sorely pressed for time.

Given that their role is primarily to influence and not to execute, a critical requirement of the ideas that they promote is that they should be expressed in terms that are already understood. Like successful advertising, the messages this group propounds have to be conveyed in the briefest of moments. This pressure contributes to an environment where it is very difficult to operate other than within conceptual frameworks and taxonomies which are highly institutionalised. To be effective this group has to focus on narrative rather than taxonomical detail and on stories involving big absolute numbers rather than relative frequencies.

Heavy reliance on these institutionalised terminologies and the assumption that everyone understands what is meant by them results in this group being prone to equate proxies with realities and can result in an institutional complacency where people are unwilling to invest the time and effort on new ways of seeing the world, however accurately these taxonomies capture current social realities.

(g) Civil Servants

Given the intense focus of Civil Servants on the preferences and thought processes of their (supposed) political masters, it is understandable why this group of actors should be the least receptive of any group to the new taxonomies arising from Big Data. Recruited mostly from the humanities rather than from quantitative disciplines, the culture of this group tends not to prioritise direct experience or the analysis of evidence from primary sources. The evaluation of the effect of newly implemented policies is contracted out to academic researchers and consultants; service delivery is outsourced to organizations that have no commercial interest in the use of Big Data and encounters with the subjects to whom public services are delivered is too infrequent to sensitise this group to whether existing taxonomies adequately capture the variety of the contemporary social world.

Among this group it is the policy agenda that drives an increasingly tightly prescribed research requirements agenda. One can picture the dinner table guests of Sir Humphrey Appleby ²⁰ – or even Lord (Gus) O’Donnell – recounting anecdotes to illustrate personal idiosyncrasies or describing accidental encounters in the social world (‘according to my Colombian cleaning lady …’), then sharing their experiences of focus groups or imagining smart ways of using administrative data to reduce the reporting burdens under which business or the public sector suffer.

If members of this group did subscribe to the view of Peters and Rogers that information management is a principal source of competitive advantage, it is questionable whether they would apply it to the process of developing public policy (partly because they do not have competitors in what they regard as a monopoly over policy). When celebrating the creative industries and scientific innovation which they claim to be a hallmark of the modern British economy, it is doubtful whether they would suppose that their own ways of conceptualising the citizenry could benefit from innovation of this sort.

Conclusion

This paper develops the general contention made by Savage and Burrows that Big Data can provide insights into social behaviour which are beyond the reach of traditional survey questionnaires. It is our belief that it is not helpful to criticise particular groups for not making greater use of Big Data than they have.

Whilst we believe that it is not possible to provide a theoretical explanation for the uneven adoption of Big Data as a whole, we believe it is possible to identify the reasons for differential adoption of the generic taxonomies that have been developed principally from Big Data sources. We contend that examining the differential adoption of this particular form of Big Data is useful since it is one where differences in adoption rates cannot be explained either by the data being proprietary and confidential in nature or by the cost of accessing it.

On the contrary, we conclude that the uneven adoption of these taxonomies is better explained by differences in the circumstances of potential user groups and the relationship that exists between the objectives and standards that guide research practice and the features which differentiate new taxonomies from traditional ones, some of which are off-putting to certain groups but attractive to others.

We conclude that the adoption of generic taxonomies based on Big Data is likely to be faster in environments where it is possible to measure improvements in the achievement of specific performance metrics. Adoption is slower where the evaluation of research work rests on its contribution to theories that are explicit and methods are the common currency of the peer group that provides the most important formal means of evaluation (through status, promotions and appointments), that have been established through a collaborative programme of research involving multiple stakeholders and in which the evaluation process is characterised by extensive, external review. On the basis of our experience generic taxonomies based on Big Data are more likely to be adopted where external stakeholders are relatively little involved in directing research activity and where, as a result, a specialist analytic team is granted a relatively high level of autonomy to implement “whatever works best” in a results-driven environment. Acceptance is likely to be greatest where there is long term continuity in the management of the analytic team and where the investment in understanding new taxonomies can confidently be deployed over a longer period than a typical research grant or ministerial appointment.

The conclusion that we have drawn from our analysis is that the adoption of taxonomies based on Big Data is far from random, that actors' reactions are much less capricious than we had supposed and that it has proved much easier to develop theoretical explanations for the uneven use of these systems than we had originally expected.