Big data in scientific research

Some scientific work requires the accumulation of huge amounts of data. In the field of astronomy, for example, it has been reckoned that the Sloan Digital Sky Survey (SDSS), based in New Mexico, accumulated more data within a few weeks of its commencing operations in 2000 than had previously been acquired in the entire history of astronomy—and that a scheduled future astronomical project will acquire every 5 days the amount of data accrued by the SDSS over the course of 10 years (Mayer‐Schönberger and Cukier , 7). Similarly, the huge amounts of information generated through scientific research programs such as gene sequencing, and the experiments being undertaken at the Large Hadron Collider at CERN (Chen et al. 2013, 21–23), have the potential, on analysis, to advance human knowledge and understanding greatly in the areas of research they are addressing.

Commercial exploitation of big data

Information about the purchasing habits of individuals and communities, and also information about the manufacturing and distribution of goods, and about the internal organization of large companies, is of tremendous commercial significance (Davenport and Dyché 2013). A great many financial transactions now routinely take place on the Internet, enabling data to be collected by those selling goods and services; and the use of store cards and credit cards in retail outlets also allows information about individuals’ purchasing habits to be compiled. Companies can use big data of this kind in an “inward‐facing” way, to streamline their own internal processes and make them more efficient, and in an “outward‐facing” way, using data retrieved from customers to analyze those customers’ preferences and hence to advertise products which those customers are likely to want to buy. Users of retail websites such as Amazon will be aware that items for which they search are routinely logged, and that the information thus supplied is used to make recommendations for future purchases. Most users of such websites will probably not find this use of the data they supply problematic, and may indeed find it helpful.

Big data in medical research

Information gleaned from hospital records, and from the results of medical trials, is potentially of great importance in the forecasting, diagnosis, and treatment of disease (Chen et al. 2013, 73–4). This makes the large data sets that may be gathered through the scrutiny of patient records extremely valuable for pursuing scientific research; and it also means that they constitute a valuable resource from the perspective of those who might wish to exploit such research for profit, such as pharmaceutical and insurance companies. There are, of course, standard ethical issues around the collection and storage of such data which pre‐date the big data age (cf. British Medical Association , esp. chapters 2 and 5); but big data introduces new issues of its own, some of which will be explored briefly below.

An example of the use of big data

At this point, a much‐commented‐upon example will perhaps serve to illustrate some of the potential benefits of big data, and some of its potential pitfalls. The Google Flu Trends (GFT) service is an attempt by the Internet search service Google to make accurate predictions about flu outbreaks (Ginsberg et al. 2009). In the United States, these outbreaks can be tracked through the analysis made by the Centers for Disease Control and Prevention (CDC) of the visits made to physicians by flu sufferers: these data are published with a lag time of 1–2 weeks. A number of common search queries were analyzed to discover which best fitted this CDC data, using archived information on Google usage. Attempts were then made to provide a much quicker analysis of the occurrence of flu epidemics, with the aim of  “enabl[ing] public health officials and health professionals to better respond to seasonal epidemics” (Ginsberg et al. 2009, 1013). Subsequent work suggested that “Google Flu Trends can provide timely and accurate estimates of the influenza activity in the United States, especially during peak activity, even in the wake of a novel form of influenza” (Cook et al. , 7).

However, a more recent U.S. flu epidemic in winter 2012–2013 was presaged by “drastically overestimated peak flu levels” predicted by GFT (Butler , 155). Although it is likely that the refinement of the methods used in making GFT predictions in the light of such wayward results will subsequently reduce their occurrence, it has been suggested (Lazer et al. ) that two issues have been particularly significant in producing these errors: big data hubris and algorithm dynamics. “‘Big data hubris’ is the often implicit assumption that big data are a substitute for, rather than a supplement to, traditional data collection and analysis. … The core challenge is that most big data that have received popular attention are not the output of instruments designed to produce valid and reliable data amenable for scientific analysis” (Lazer et al. : 1203). In addition to such hubris, the more practical point is made that Google constantly modifies the algorithms it is using, in order to try to improve its service to its customers (for example, by making recommendations for further searches) and in response to its business model (for example, by promoting advertising revenue). Such algorithm modification has the potential to introduce instability into predictive systems such as GFT. Lazer et al. () suggest that big data offers significant possibilities, but that there remain issues around transparency (those working on GFT do not give the search terms they identified as most significant), understanding how algorithms change over time, and integrating research on big data and the more traditional “small data”—that is to say, data collected in traditional ways, by repeatable means.

In the work summarized here by those engaged with GFT, and by those offering a critique of their approach, we may perceive the beginnings of a critical interrogation of big data. This is a topic to which we shall return in the section on hermeneutics.

Ethical challenges (i): Confidentiality and consent

In both the generation and storage of big data sets containing personal information, there are issues raised around the confidentiality of the information involved. “Confidentiality” here involves preserving information entrusted to the generator of the data set from access by an unauthorized third party.

To some extent, it may be possible to preserve confidentiality through processes of anonymization or “deidentification”—that is to say, “stripping information from a data record that might link the record to the public name of the record's subject” (Berman : 28). Such a process would need to be irreversible if confidentiality is to be maintained (which may then limit the uses of the data). However, it has been suggested that given enough data, anonymization is all but impossible, as it is possible to triangulate from multiple “anonymized” data sets and hence re‐establish individual identities (Mayer‐Schönberger and Cukier , 154ff). Hence the rise of big data raises new issues around the practicability of maintaining individual confidentiality. (We may note in passing that in the case of the GFT research cited above, the researchers state that “None of the queries in the Google database for this project can be associated with any particular individual” (Ginsberg et al. 2009, 1014).)

Similarly, new issues are raised by big data around the obtaining of informed consent for research. This is a routine procedure in the conduct in all kinds of research, medical, psychological, sociological, and so on, involving human subjects (cf. British Medical Association , 59–63). There are straightforward practical problems here, since the generation and storage of such consent forms is itself a big data issue, and so the questions already raised around confidentiality apply to them; also, obtaining consent can be a very expensive process in terms of the time involved (Berman , 190). There is also a more theoretical problem that arises with the storage of data which may then be open to revisiting for future research. To what exactly are people consenting? Suppose I give a blood sample as part of a research program seeking information about susceptibility to a particular infectious disease. If the information gleaned from that sample is retained as part of a large database, and is subsequently used in research relating to cancer, should my consent be re‐sought for this new research? In general, if data is being retained for (potentially) multiple uses, some of which may not be known at the time when consent was obtained, to what extent can the consent given be truly “informed?” Similarly, if blanket consent were to be sought from individuals for potentially any application of data derived from their submissions to a research project, to what extent could consent given under such circumstances be said to be “informed?”

It has been suggested that a “privacy framework for the big data age” will be “one focused less on individual consent at the time of collection and more on holding data users accountable for what they do” (Mayer‐Schönberger and Cukier , 173). But this itself raises a new set of questions, not least regarding how boundaries for accountability may be set, and how legislation might be enacted on an international basis to ensure that such accountability could be maintained when data from the population of more than one nation‐state is involved. Considerably more work is needed in the areas of confidentiality and consent if the new challenges presented by big data are to be fully addressed.

Ethical challenges (ii): Storage and access

“Over the past half‐century, the cost of digital storage has been roughly cut in half every 2 years, while the storage density has increased 50 million‐fold” (Mayer‐Schönberger and Cukier , 101). It is to be expected that these trends—reducing expense and increasing capacity—will continue, with the result that more and more data will be accumulated and stored. This presents at least two potential problems concerning the secure storage of big data: first, ensuring that data are neither accessed without authorization nor stolen; and second, ensuring that it remains uncorrupted. The widely reported hacking of Sony's computer systems in December 2014 and the consequent revelation of confidential information on the Internet, in a supposed “revenge” attack over its proposed release of a movie depicting the assassination of the North Korean leader Kim Jong‐un, illustrates how real the first of these dangers is, and anyone who has been the victim of a computer virus will realize how vulnerable computer systems can be to corruption and degradation. The secure storage of data is an ongoing problem, and is likely to remain one indefinitely as both security systems and the ingenuity of hackers become more sophisticated.

Leaving aside questions arising from real or potential criminal activity, it has been said that “the most important question confronting the future of Big Data [is]: will it be open (to the public) or closed (to all but the data owners)?” (Berman , 224). Given the variety of possible readings that may be generated by big data (see below), it is important that that data be readily available to anyone who might wish to interrogate it, not least in order to confirm or challenge conclusions which have been derived from it. For a variety of reasons, Berman is pessimistic in his appraisal of this issue, concluding that for commercial, practical, and propagandistic reasons there are vested interests involved in restricting access to big data (Berman , 224), to the extent that the kind of free access that is desirable is unlikely to materialize in reality. This in turn creates a raft of questions around the credibility of ideas advanced on the basis of big data, if the data in question are not available for scrutiny. (It is, after all, axiomatic that scientific experiments should be repeatable: to publish results without accounting in detail for their provenance would, in classical scientific contexts, be impermissible.)

In all the areas discussed above, there is an opportunity for specifically Christian ethical engagement with the challenges brought by big data. There are important opportunities here for dialogue between theologians and those involved with the management of big data. Such dialogue may not be straightforward, and where business issues are concerned it may involve a re‐examination of the roots of contemporary economic life (cf. Stackhouse ). This is an important area for future research.

Interpretative challenges: Hermeneutics

Perhaps the most significant of all the challenges relating to big data are those that cluster around its interpretation. Very big data sets can offer support for any number of responses to any number of questions used to interrogate it, as Berman (, 145) points out: “When the amount of data is sufficiently large, you can find almost anything you seek lurking somewhere within. … Also, whenever you select a subset of data from an enormous collection, you may have no way of knowing the relevance of the data you excluded.” He adds that “Data scientists walk a thin line. If they start their project with a preconceived theory, then they run the risk of choosing a data set that confirms their bias. If they start their project without a theory, then they run the risk of developing a false hypothesis that happens to fit the data” (Berman , 147). Berman surveys various pitfalls of big data analysis, including the biases that the analytical techniques used can themselves introduce. His conclusion should give us pause for thought: “Technically, Big Data does not produce answers. At best, Big Data points us in the direction of answers and inspires new questions. At worst, Big Data pretends to give us answers when it cannot” (Berman , 226).

We will return to the important issue of interpretation in the sections on the data scientist and hermeneutics below.

An interdisciplinary exercise

In order to maximize the benefit that may be obtained from the analysis of big data, it is important to have a number of different perspectives on it. In particular, data originating in particular scientific, clinical, or commercial contexts is very likely to require the involvement of appropriate subject specialists in its interpretation. It may well be that a simple consideration of the statistics throws up odd correlations, the interpretation of which requires an expert eye. (In an amusing case reported by Mayer‐Schönberger and Cukier [, 67], statisticians commissioned by a used‐car trader found that cars painted orange had fewer engine defects than those painted other colors. It is not immediately obvious what interpretation might be placed on this correlation; in generating any such interpretation, expertise from the manufacturers and distributors of cars is likely to be necessary.)

Issues of meaning

Just as meaning emerges from an interaction of reader and text, likewise meaning emerges from the interaction of data and an analyst. In both cases, it is important to pay close attention to the meaning of that meaning. Texts tell us stories: so too do data. Any meanings derived from the presentation of those stories—the “messages” or “morals” they are presenting—require careful and scrupulous consideration. This is particularly the case when it comes to the telling of stories derived from big data, which are likely to be accompanied by graphics—charts, graphs, and diagrams—which may themselves embody, in an impressive or even apparently irrefutable form, biases or misreadings of big data which are consequent on the method of its analysis.

The practice of data analysis

As already noted, the treatment of data requires skill. Understanding, interpreting, and presenting data to others will require future data scientists to be trained, and to practice their craft in the expectation that their skills, and the wisdom they bring to bear in their work, will increase with experience. (“Wisdom” here may usefully be understood as “an interpretation of knowledge that is not separated from the ethical claims of truth and goodness” [Deane‐Drummond , 153]). Not only this: as Michael Polanyi commented, “Making sense of experience is a skillful act, which impresses the personal participation of the scientist on the resulting knowledge” (Polanyi , 60). It is important to recognize that as in the conduct of experiments, so in the analysis of data, the individual scientist brings a unique perspective to bear, and leaves in consequence a unique personal imprint on the results he or she generates.

The context of data

Just as it is important to have some understanding of the original historical, cultural, and social context of a text if any interpretation of it is to be of value, so it is important to understand as fully as possible the provenance of any data that are being used if maximum benefit is to be obtained from them. Many phenomena have a complex etiology and cannot be fully understood unless as wide a picture of them as possible is obtained. Contributory factors to a particular disease, for example, may include genetic, familial, environmental, economic, and other factors, not all of which may be revealed through an analysis based solely on hospital records.

The bias of the analyst

“Interpretation has some specific subjective connotations, such as the reciprocity between interpretation of the text and self‐interpretation” (Ricoeur , 26). Interpreters of a text need to interrogate their own historical, cultural, and social contexts, and the expectations that these place on them, in addressing their task. So, too, the interpreters of data need to interrogate the expectations and demands that their contexts place on them. This does not include just their social and cultural contexts, but may well involve such questions as: do future promotions, or salary rises, hinge on my deriving the “right” results from this data? Am I using the data I am analyzing in order to seek some kind of competitive edge in the field in which I am operating, either personally or on behalf of my employer?

The role of community

The meanings of texts are shaped by communities: so too are the meanings of big data. The attitudes, practices, and beliefs of the “data science community,” and of any corporate community which data scientists are serving, will inevitably affect the way in which those data scientists operate. This may have a positive role, in reinforcing good practice; or it may have a more negative role, in discouraging original thinking. Much data is currently treated as “junk,” and discarded; for example, it is estimated that less than 0.1% of the data generated by the Large Hadron Collider at CERN is currently collected and stored (Mayer‐Schönberger and Cukier , 197). This information is presumably, by a general consensus, reckoned to be the most useful; but who knows what potentially valuable information, currently seen as worthless, might turn out to have been lost as a consequence of this consensus among the community of practitioners working at CERN?

The rejection of unique, objective meanings

To some extent, one might hope to be preaching to the choir on this issue, as far as data scientists are concerned: as we have seen, there is already a realization that any number of meanings may be derived from big data sets. However, given the particular status which is generally accorded to mathematized conclusions, which may well appear to be “objective” in the ways in which they are presented, it is perhaps also incumbent on us to remember at all times that big data can be fallible: that not all the correlations it presents to us will mean anything.

In the case cited earlier in this article—the use of big data by GFT—and in the subsequent analyses of that project's shortcomings, it may be seen that many of these hermeneutical principles are starting to be applied. The work of GFT is interdisciplinary, involving computer scientists and epidemiologists. There is a recognition in the critiques of GFT analyses that any “meaning” derived from the data used requires scrutiny and checking; and a realization, too, that big data originates from a particular context, and is shaped by the algorithms used to collect it. Concerning potential analyst bias, it is noteworthy that Cook et al. () explicitly acknowledge the “competing interests” in their research, in that three of the authors are Google employees. Finally, we may perhaps also see here some evolution of the “big data community” as those engaged in big data research initiate and respond to a particular instance of its use, seeing it as “a case study where we can learn critical lessons as we move forward in the age of big data analysis” (Lazer et al. , 1205).